Cervical cell panorama-oriented multi-round dialogue type intelligent diagnosis interaction system

Abstract

The invention relates to a multi-round dialogue type intelligent diagnosis interaction system, in particular to a cervical cell panorama-oriented multi-round dialogue type intelligent diagnosis interaction system. The objective of the invention is to solve the problems that existing diagnostic information is single, insufficient in explanatory property and poor in interaction flexibility, and finer diagnostic information is difficult to obtain; the problems that multi-round interaction and context memory ability are lacked, and continuous and coherent response cannot be made in combination with historical information are solved. Comprising an image block acquisition module used for acquiring a pre-processed full slice image WSI; the image segmentation and effective area screening module is used for obtaining effective image blocks; the image-text pair construction and pathological attribute acquisition module is used for forming attribute vectors; the context-aware multi-round question and answer module is used for obtaining a final answer; and the final diagnosis suggestion generation module is used for obtaining a final diagnosis suggestion based on knowledge graph reasoning and probability model prediction. The method is applied to the field of multi-round dialogue type intelligent diagnosis interaction.

Description

Technical Field

[0001] This invention relates to a multi-turn conversational intelligent diagnostic interactive system. Background Technology

[0002] Cervical cancer, as one of the most common malignant tumors in women, has become a significant public health issue for women worldwide. For a long time, cervical cancer screening has primarily relied on cytological smears under a manual microscope, namely the traditional Pap smear technique. However, this traditional method is highly dependent on the experience and expertise of pathologists, and is easily influenced by human factors, leading to significant subjectivity and instability in diagnostic results.

[0003] With the development of digital pathology technology and scanner equipment, high-resolution whole-slide images (WSI) have gradually become an important technical means for modern pathological diagnosis. WSI images have extremely high resolution, often containing tens of thousands or even hundreds of thousands of cells in a single image. Furthermore, the cellular and tissue structures are complex, and abnormal cells are often sparsely distributed and concealed. Doctors face a significant challenge in accurately identifying all abnormal areas within a limited time using traditional manual image reading methods, resulting in high workload and a high risk of missed diagnoses.

[0004] To address the aforementioned issues, numerous AI-based assisted diagnostic systems have emerged in recent years, attempting to automate the localization and classification of abnormal regions through deep learning-based classification, semantic segmentation, or detection models. These solutions typically assist doctors in quickly locating lesions by providing fixed-category diagnostic labels or segmentation masks. However, traditional AI systems generally suffer from a "black box problem," meaning that the diagnostic labels or masks output by the model are difficult to explain to doctors in detail, lacking a clear reasoning process. This lack of interpretable output prevents doctors from intuitively understanding the internal logic and basis of the AI ​​system's diagnosis, making it difficult to develop trust and reliance on AI.

[0005] Furthermore, current AI-assisted diagnostic systems are almost entirely one-way information output systems, meaning the model only outputs fixed diagnostic results, and doctors cannot effectively interact with the model. Especially in clinical practice, doctors often need to inquire about more in-depth information regarding specific abnormal areas, such as a detailed description of the morphological characteristics of cells in a certain area, the reasoning behind the degree of abnormality, and comparative analysis of similar historical cases. Existing systems cannot provide this dynamic and flexible information interaction mechanism.

[0006] Therefore, the existing technology still has the following shortcomings:

[0007] (1) Diagnostic information is limited and lacks interpretability: The model usually outputs results in the form of fixed labels or masks, which cannot explain the logical basis and reasoning path of the diagnosis;

[0008] (2) Poor interaction flexibility: Doctors cannot have in-depth interactive dialogues with the model through natural language, making it difficult to obtain more detailed diagnostic information;

[0009] (3) Lack of multi-turn interaction and context memory: In continuous question-answering scenarios, existing models cannot combine historical information to make continuous and coherent responses.

[0010] In summary, existing AI-assisted systems for cervical cytology diagnosis cannot yet meet the clinical demands for interpretability, interactivity, and consistency in model output. Therefore, there is an urgent need to develop an intelligent diagnostic system that fully integrates image features with natural language interaction capabilities, using an interactive question-and-answer approach to help doctors quickly, accurately, and reliably complete the diagnosis of cervical cytology. Summary of the Invention

[0011] The purpose of this invention is to address the problems of existing diagnostic information being singular and lacking in interpretability, having poor interactive flexibility, making it difficult to obtain more detailed diagnostic information, lacking multi-turn interaction and contextual memory capabilities, and being unable to make continuous and coherent responses by combining historical information. Therefore, this invention proposes a multi-turn conversational intelligent diagnostic interactive system for cervical cell panorama.

[0012] A multi-turn conversational intelligent diagnostic interactive system for cervical cell panorama includes:

[0013] The module includes: image patch acquisition module, image segmentation and effective region filtering module, image-text pair construction and pathological attribute acquisition module, context-aware multi-turn question answering module, and final diagnosis suggestion generation module.

[0014] The image block acquisition module is used to acquire the whole slice image WSI, and to preprocess the acquired whole slice image WSI to obtain the preprocessed whole slice image WSI.

[0015] The image segmentation and effective region filtering module is used to divide the preprocessed whole slice image WSI into blocks to obtain image blocks, and filter the image blocks to obtain effective image blocks.

[0016] The image-text pair construction and pathological attribute acquisition module is used to construct instructions (Prompt) for the obtained valid image patches. The constructed instructions (Prompt) and valid image patches are input into the Visual Large Language Model (VLM), and the Visual Large Language Model (VLM) outputs text. Pathological attributes are parsed and quantified from the text to form attribute vectors.

[0017] Context-aware multi-turn question answering modules are used for fusion context representation And the optimal instruction Prompt yields the final answer;

[0018] The final diagnostic suggestion generation module is used to calculate the average anomaly score based on the attribute vector, perform knowledge graph reasoning based on the average anomaly score, perform probabilistic model prediction based on the attribute vector, and obtain the final diagnostic suggestion based on knowledge graph reasoning and probabilistic model prediction.

[0019] The beneficial effects of this invention are as follows:

[0020] To address the shortcomings mentioned in the background technology, this invention proposes an innovative interactive system for cervical cell WSI diagnosis based on a multi-turn natural language question-answering mechanism. It comprehensively utilizes a large visual-language model, contextual understanding technology, natural language generation technology, and knowledge-enhanced reasoning mechanisms to solve the following key problems:

[0021] Achieving natural language interaction between doctors and AI systems: Traditional AI-assisted diagnostic systems only output fixed results, preventing doctors from engaging in in-depth interaction. This invention aims to construct a human-computer interactive dialogue mechanism that enables doctors to ask questions flexibly using natural language, such as "Where is the abnormal area?" or "Are there any cellular pathological features?", thereby effectively improving the comprehensibility and information depth of the diagnosis.

[0022] This invention achieves automatic identification and detailed textual description of abnormal cell regions without relying on traditional image segmentation or classification. Current methods generally rely on labeled data for supervised learning, resulting in singular and opaque outputs. This invention, through fine-tuning of a vision-language model and Prompt engineering design, automatically generates precise descriptions of cellular pathological features (such as abnormal nucleocytoplasmic ratio, chromatin abnormalities, etc.), automatically identifying lesion areas and outputting detailed textual descriptions without the need for traditional segmentation labels.

[0023] This invention introduces a multi-turn question-and-answer contextual memory and reasoning mechanism to ensure the coherence and rationality of responses: Doctors often ask a series of questions during clinical diagnosis, but traditional AI systems lack historical question-and-answer memory and cannot respond coherently. This invention utilizes advanced memory enhancement mechanisms (such as a RAG-based knowledge-enhanced memory module and historical dialogue embedding technology) to semantically understand and continuously track the historical question-and-answer context, ensuring logical coherence and consistency in the question-and-answer content.

[0024] Enhancing the transparency and credibility of model diagnostic results to ensure they better align with clinical diagnostic logic: Existing AI systems struggle to explain diagnostic causes, significantly limiting their clinical application. This invention constructs an interpretable reasoning mechanism, employing multimodal feature fusion and explicit reasoning chain generation technology to ensure the diagnostic process is transparent and visible, with results possessing a clear medical reasoning path, thereby increasing doctors' trust in the model's outcomes.

[0025] Ultimately, this invention aims to provide efficient, accurate, and reliable cervical cytology diagnostic assistance, effectively reducing the burden on clinicians in reading images. The interactive diagnostic system of this invention can not only automatically and quickly locate and describe the diseased cell areas, but also provide highly structured and intuitive diagnostic suggestions and risk warnings in natural language. This enables doctors to quickly and effectively obtain key information during image reading, greatly improving diagnostic efficiency and accuracy, and reducing the burden and pressure on clinicians in reading images.

[0026] In summary, this invention provides an innovative image-text fusion diagnostic interaction framework that, through interactive question-and-answer and natural language interpretation, fully meets the urgent clinical needs for AI-assisted diagnostic systems in terms of interpretability, interactive flexibility, and credibility, and significantly promotes the development of AI technology for cervical cytology diagnosis towards practicality and intelligence. Attached Figure Description

[0027] Figure 1 This is a diagram of the overall architecture of the present invention; Figure 2 To create a flowchart of the process and processing steps; Figure 3 This is a schematic diagram of the core modules and technologies; Figure 4 This is a schematic diagram of the workbench of the intelligent diagnostic system; Figure 5 A schematic diagram of a human-computer collaborative dialogue module; Figure 6 This is a schematic diagram of an automatically generated diagnostic report. Detailed Implementation

[0028] Specific Implementation Method 1: This implementation method's multi-turn dialogue-based intelligent diagnostic interactive system for cervical cell panoramic imaging includes:

[0029] The module includes: image patch acquisition module, image segmentation and effective region filtering module, image-text pair construction and pathological attribute acquisition module, context-aware multi-turn question answering module, and final diagnosis suggestion generation module.

[0030] The image block acquisition module is used to acquire a high-resolution full-slice image (WSI), and to preprocess the acquired high-resolution full-slice image (WSI) to obtain a preprocessed full-slice image (WSI).

[0031] The image segmentation and effective region filtering module is used to divide the preprocessed whole slice image WSI into blocks to obtain image blocks, and filter the image blocks to obtain effective image blocks.

[0032] The image-text pair construction and pathological attribute acquisition module is used to construct instructions (Prompt) for the obtained valid image patches. The constructed instructions (Prompt) and valid image patches are input into the Visual Large Language Model (VLM), and the Visual Large Language Model (VLM) outputs text. Key pathological attributes are parsed and quantified from the text to form attribute vectors.

[0033] Context-aware multi-turn question answering modules are used for fusion context representation And the optimal instruction Prompt yields the final answer;

[0034] The final diagnostic suggestion generation module is used to calculate the average anomaly score based on the attribute vector, perform knowledge graph reasoning based on the average anomaly score, perform probabilistic model prediction based on the attribute vector, and obtain the final diagnostic suggestion based on knowledge graph reasoning and probabilistic model prediction.

[0035] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that: the image block acquisition module is used to acquire a high-resolution full-slice image WSI, and preprocesses the acquired high-resolution full-slice image WSI to obtain a preprocessed full-slice image WSI; the specific process is as follows:

[0036] S11. Use the OpenSlide open-source library to obtain the high-resolution full-slice image WSI and the metadata of the full-slice image WSI;

[0037] Metadata for whole-slice images (WSI) includes: the physical size of the WSI image (micrometers / pixel), the scan magnification (e.g., 20x, 40x), and the resolution of each level in the image pyramid; this metadata is the basis for achieving subsequent accurate quantitative analysis (such as cell nuclear area calculation) and ensuring consistency in multi-scale processing.

[0038] The OpenSlide open-source library is used to ensure broad compatibility with WSI formats generated by various mainstream digital pathology scanners, such as Aperio SVS, Leica SCN, Hamamatsu NDPI, and standard TIFF.

[0039] WSI is essentially a collection of multi-resolution images. Full-slice WSI images are explicitly constructed as an image pyramid structure:

[0040]

[0041] in, WSI for full slice images; It provides rich details of the cell nuclei for the highest resolution (e.g., 40x objective) images in WSI for whole-slice images; This is the second highest resolution image (e.g., 40x objective) in the whole-slice image WSI; This is the lowest resolution thumbnail in the WSI full-slice image, used for a quick overview of tissue distribution;

[0042] This pyramid structure is crucial because it allows the system to perform analysis at different scales: rapidly locating regions of interest at low-resolution levels and performing precise cytological assessments at high-resolution levels, perfectly mimicking the operation of a pathologist switching magnifications under a microscope.

[0043] S12. Preprocess the acquired high-resolution whole-slice image WSI to obtain the preprocessed whole-slice image WSI; the specific process is as follows:

[0044] S121. Use a standard color normalization algorithm (such as the method of Macenko et al.) to normalize the lowest resolution thumbnail in the whole slice image WSI. Correction is performed to obtain the corrected high-resolution whole-slice image WSI;

[0045] S122. Input the corrected high-resolution whole-slice image WSI into the U-Net segmentation model. The U-Net segmentation model outputs a binarized tissue region mask. ;

[0046] Binarized tissue region mask The corrected high-resolution whole-slice image WSI indicates a blank background area; the binarized tissue region mask. The corrected high-resolution whole-slice image WSI indicates the cervical cell region;

[0047] Preservation of binarized tissue region mask Corrected high-resolution whole-slice image WSI;

[0048] Color Normalization: Due to differences in staining agents, scanners, and operating procedures between different laboratories, WSI images often exhibit significant color shifts. To eliminate the interference of this "batch effect" on model performance, the system will use standard color normalization algorithms (such as the method of Macenko et al.) to correct the images, making their color distribution more consistent.

[0049] Tissue Mask Generation: WSI contains a large number of blank background areas, and processing them directly would result in a huge waste of computational resources. This step utilizes low-resolution images. By combining Otsu thresholding, color space analysis (such as HSV), or a lightweight U-Net segmentation model, binarized tissue region masks can be generated quickly and accurately. This mask will serve as a "map" for all subsequent spatial operations, ensuring that analysis is performed only within meaningful tissue areas. Other steps and parameters are the same as in Specific Implementation Method 1.

[0050] Specific Implementation Method 3: This implementation method differs from Specific Implementation Method 1 or 2 in that: the image segmentation and effective region filtering module is used to divide the preprocessed full slice image WSI into blocks to obtain image blocks, and to filter the image blocks to obtain effective image blocks;

[0051] The goal of this stage is to decompose the macroscopic WSI into information-intensive micro-units (Patches) that the model can process;

[0052] S21. Divide the preprocessed full-slice image WSI from S12 into blocks to obtain image patches; the specific process is as follows:

[0053] In the image pyramid structure, select an intermediate layer image that balances cellular details and local context (such as Level 1, corresponding to...). or (Objective field of view) ;

[0054] The overlapping sliding window strategy is used to separate the intermediate layer images. Cut into a large number of fixed-size image patches:

[0055]

[0056] in, To make the intermediate layer image A collection of image blocks cut into fixed sizes; For set The first image patch; For set The second image patch in the middle; For set Middle Image blocks; For set Middle Image blocks; To make the intermediate layer image The total number of image blocks cut into fixed sizes; To make the intermediate layer image The first image block cut into fixed size indivual; ; , For real numbers, For the first Image blocks of high, For the first Image blocks The width, 3 is the first Image blocks RGB3 channels;

[0057] Setting a certain overlap rate (such as 25%) can effectively prevent cells from being cut at the cutting boundary, ensuring that each cell can appear intact in the central area of ​​at least one patch.

[0058] S22. Filter the image blocks to obtain valid image blocks; to further refine the data, process each image block... A double screening process is performed; the specific process is as follows:

[0059] S221, the set of image patches obtained in S21 Calculate tissue coverage for each image patch ;

[0060] Set tissue coverage threshold , ;

[0061] Eliminate Image blocks, preserved Image blocks;

[0062] Eliminate The background is the main image block to filter out the glass plate and gel areas;

[0063] The image patch set obtained from S21 Calculate tissue coverage for each image patch The expression is:

[0064]

[0065] in, Indicates the first Image blocks The corresponding mask matrix (with values ​​of 0 or 1); This represents the pixel threshold, with a value of 0.5. Represents image blocks The mask corresponding to each pixel is greater than the pixel threshold. The number of image blocks; Indicates the first Image blocks Organizational coverage;

[0066] S222, Reservation of S221 Calculate the information entropy for each image patch; set the information entropy threshold;

[0067] Image blocks with information entropy less than the information entropy threshold are discarded as valid image blocks;

[0068] Information entropy filtering: For patches that pass the coverage filtering, calculate their information entropy. Patches with low information entropy usually correspond to blurry, out-of-focus, or monotonous areas and should be removed, thus focusing computational resources on high-quality, high-information areas. Other steps and parameters are the same as in specific implementation methods one or two.

[0069] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that: the image-text pair construction and pathological attribute acquisition module is used to construct a Prompt instruction for the obtained valid image blocks. The constructed Prompt instruction and the valid image blocks are input into the Visual Large Language Model (VLM), and the VLM outputs text. Key pathological attributes are parsed and quantified from the text to form attribute vectors. The specific process is as follows:

[0070] S31. Construct a Prompt instruction for each valid image block obtained in S22 (the Prompt instructions for all valid image blocks are the same); the number of Prompt instructions corresponding to each valid image block is greater than or equal to 1 and less than or equal to 5.

[0071] The Prompt command is:

[0072] Basic Description: Please describe the first part in detail. A valid image patch (a specific one within a valid image) includes: First, evaluating the... The quality and cell density of each effective image patch were then analyzed; subsequently, the main first-order image patches were identified and described. Cell types of the first effective image patch; finally, pay special attention to the first... Does the valid image block contain abnormal cell features?

[0073] Feature query: Do each valid image patch have cells with the following characteristics: 1) enlarged and irregular nuclei; 2) significantly increased nucleocytoplasmic ratio; 3) coarse granular or unevenly distributed chromatin; 4) presence of deep nuclear staining?

[0074] For each filtered valid Instead of using simple category labels, it constructs a text-image input pair to guide the model in logical reasoning. Its core is a structured task instruction prompt, which draws inspiration from the Chain-of-Thought (CoT) concept to guide the model through step-by-step analysis.

[0075] S32, the first obtained in S22 The Prompt instruction for the first valid image block and the... Each valid image patch is input into the Visual Large Language Model (VLM), which outputs text. Key pathological attributes are parsed and quantified from the text to form attribute vectors. The specific process is as follows:

[0076] S321. The Visual Large Language Model (VLM) is fine-tuned using LoRA technology to obtain the fine-tuned VLM. A powerful pre-trained multimodal large model (such as LLaVA-1.5, MiniGPT-4, etc.) is selected, and LoRA (Low-Rank Adaptation) technology is used for efficient domain-specific parameter tuning. LoRA can inject cervical cytopathology expertise without changing most of the weights of the original model. The specific process is as follows:

[0077] S3211, The forward propagation process of a certain layer of the Visual Large Language Model (VLM) becomes:

[0078]

[0079] in, It is a low-rank matrix. ; It is a low-rank matrix. ; The token vector is the input to a certain layer of the visual large language model; It is a scaling constant; This is the weight matrix of the visual large language model; For rank; The dimension of the token vector input to a certain layer; The dimension of the token vector output by a certain layer; This is the output vector after incorporating the low-rank LoRA update.

[0080] S3212. Constructing a multi-task joint loss function for a visual large language model (VLM). Multi-task joint loss function Convergence yields the fine-tuned Visual Large Language Model (VLM).

[0081] Multi-task joint loss function for:

[0082]

[0083] in, This represents the text generation loss, which is the cross-entropy loss. express The weight hyperparameters; Indicates classification loss; express The weight hyperparameters; This represents the regression loss, which is the mean squared error loss. express The weight hyperparameters;

[0084] S322, the first obtained in S22 The Prompt instruction for the first valid image block and the... The fine-tuned Visual Large Language Model (VLM) takes n valid image patches as input and outputs the nth valid image patch. Each image and text pair corresponds to a text. ;

[0085]

[0086] in, This represents the fine-tuned Visual Large Language Model (VLM). A coherent natural language description is represented as: the first... In the first effective image patch, the cell distribution is sparse, with a small number of squamous epithelial cells visible; the second... In one valid image block, a cell cluster (located in the center of the image) shows a significant increase in nucleocytoplasmic ratio and deep nuclear staining, with coarse granular chromatin, resembling the characteristics of high-grade squamous intraepithelial lesion (HSIL).

[0087] S323. Parse and quantify key pathological attributes from the text to form attribute vectors; the specific process is as follows:

[0088] To facilitate subsequent quantitative analysis and aggregation, from the text Key pathological attributes are obtained, forming attribute vectors. ; indicates as:

[0089]

[0090] in, This represents the probability of Koilocytosis (a type of nuclear malformation). ; The Coarse Chromatin score indicates the coarse granulation of chromatin. ; This indicates a high N / C ratio score. ; This indicates the Hyperchromasia score, which represents nuclear hyperchromatography. ; Indicates the nuclear irregularity score. ; This represents the probability of a vacuolated / hollowed-out cell (Perinuclear Halo). ; This represents the probability of multinucleation. ; Indicating an Inflammatory Background score. Each dimension corresponds to the confidence probability or score of a pathological indicator.

[0091] Attribute vector The dimensions are fixed at 8, covering the most commonly used cytological abnormalities in the Bethesda system and clinical practice. Attribute vectors Each component represents the confidence or severity of the feature's presence in the patch.

[0092] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0093] Specific Implementation Method Five: This implementation method differs from one of Specific Implementation Methods One to Four in that: the text... Key pathological attributes are obtained, forming attribute vectors. The specific process is as follows:

[0094] 1) From text based on template matching rules Key pathological attributes are obtained, forming attribute vectors. ;

[0095] 2) Based on few-shot QA (extraction for VLM / LLM) templates from text Key pathological attributes are obtained, forming attribute vectors. ;

[0096] 3) The attribute vector formed in step 1) based on template matching rules and 2) the attribute vector formed by sampling few-shot QA By performing rule-based fusion, the final attribute vector is obtained. The specific process is as follows:

[0097]

[0098] in, Represents the weight vector. It is recommended to learn on the validation set. .

[0099] The other steps and parameters are the same as those in one of the specific implementation methods one to four.

[0100] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One to Five in that: 1) based on template matching rules from text... Key pathological attributes are obtained, forming attribute vectors. The specific process is as follows:

[0101] Template matching rules include keywords | regular expression dictionary (Simplified Chinese + common synonyms), negation | uncertain trigger words, and scoring rules;

[0102] 11) Keywords | Regular expression dictionary (Simplified Chinese + common synonyms); can be expanded as needed; English terminology is only used to cover English report scenarios; the specific process is as follows:

[0103] Probability of heteronuclear Koilocytosis The keywords are: heterogeneous nucleus | heterogeneous nucleus | koilocyt | hollowed-out nucleus | perinuclear halo cell; where "|" indicates "or";

[0104] Coarse Chromatin Score The keywords are: coarse chromatin | coarse chromatin | large chromatin | uneven chromatin distribution | chromatin clumps / coarse chromatin | clumped chromatin; where " / " is the separator between Chinese and English keywords;

[0105] Here, coarse chromatin represents coarse nuclear chromatin; clumped chromatin represents clustered chromatin.

[0106] High N / C Ratio Score The keywords are: increased nucleocytoplasmic ratio | elevated nucleocytoplasmic ratio | significantly increased nucleocytoplasmic ratio | high nucleocytoplasmic ratio / high N / C; where high N / C indicates a high nucleocytoplasmic ratio; the regular expression dictionary value suggestion is: regular expression capture. ;

[0107] Hyperchromasia score The keywords are: nuclear hyperchromium | nuclear staining intensification | nuclear overstaining / hyperchrom; where hyperchromium refers to nuclear hyperchromium.

[0108] Nuclear Irregularity Score The keywords are: irregular nuclear shape | irregular nuclear outline | irregular nuclear membrane | irregular nuclear notch / irregular nuclear; where irregular nuclear indicates an irregular nuclear shape;

[0109] Perinuclear Halo probability of vacuolated / hollow cells The keywords are: perinuclear halo | vacuolated cells | vacuolation | halo; where halo represents perinuclear halo;

[0110] Multinucleation probability The keywords are: multinucleat, multinucleated giant cell, binucleat, trinucleat, multinucleatization, and multinucleat. Here, "multinucleat" indicates a multinucleated cell, and "multinucleat" indicates a multinucleated cell.

[0111] Background inflammation (Inflammatory Background Score) The keywords are: inflammatory background | inflammatory cells | neutrophilia | lymphocytosis | purulent background | inflammatory exudate / inflamm; where inflamm represents inflammation;

[0112] 12) Negation | Indeterminate trigger words; the specific process is as follows:

[0113] The trigger words for negation are: not seen | no obvious | not found | not supported | does not conform | not considered | excluded;

[0114] Uncertain trigger words include: suspicious, pending, tendency, possible, consider, mild;

[0115] 13) Scoring rules (strong / medium / weak → score); the specific process is as follows:

[0116] Strong positive keywords: significant | obvious | high | typical | large amount | severe; strong positive keywords correspond to a score of 0.9; Moderate keywords: moderate | relatively high | multiple | numerous; moderate keywords correspond to a score of 0.7; Mild keywords: mild | few | focal | occasional; mild keywords correspond to a score of 0.5; Only mentioning words without intensity; only mentioning words without intensity corresponds to a score of 0.6; uncertain trigger words correspond to a score based on the above scores. The score corresponding to the negation trigger word is 0.0.

[0117] High N / C Ratio Score Numerical hint: When At that time, the high nucleo-cytoplasmic ratio corresponds to a fraction of 0.95; when At that time, the high nucleo-cytoplasmic ratio corresponds to a fraction of 0.85; when At that time, the high nucleo-cytoplasmic ratio corresponds to a fraction of 0.75;

[0118] 14) Rule evaluation process; the specific process is as follows:

[0119] 141) Regarding the text Preprocessing is performed to obtain the preprocessed text. The specific process is as follows:

[0120] Text Each word in the text is converted to half-width (because the text...). Each English character occupies half a character width, resulting in the text converted to half-width characters. ; Convert the text to half-width characters Standardize the terminology of synonyms in the text; unify the terminology in the text. To unify the capitalization of English text; the text after unifying the capitalization of English text. Denoising (spaces, garbled text, repeated punctuation, formatting marks, table numbering, HTML language symbols) ), to obtain the preprocessed text ;

[0121] 142) For the preprocessed text The scoring process is as follows:

[0122] 1421) For the preprocessed text First, check for negation. If a negation trigger word exists, set the corresponding word to 0.0 and execute step 1422); if no negation trigger word exists, execute step 1422.

[0123] 1422) Regarding the text Detection of high N / C ratio;

[0124] If a high N / C ratio exists, the score is based on the high N / C ratio. The numerical prompt will take the corresponding score, and execute 1423); if there is no High N / C Ratio, the High N / C Ratio score will be used. The value is 0.0, so execute 1423);

[0125] 1423) Assign corresponding scores to the results of 1422) according to the keywords of strong positive / moderate / mild, and then perform 1424); the specific process is as follows:

[0126] If a strong positive keyword appears, the corresponding score is taken; if multiple strong positive keywords appear simultaneously, the maximum score is taken; if no strong positive keyword appears, the score is 0; if a medium keyword appears, the corresponding score is taken; if multiple medium keywords appear simultaneously, the maximum score is taken; if no medium keyword appears, the score is 0; if a mild keyword appears, the corresponding score is taken; if multiple mild keywords appear simultaneously, the maximum score is taken; if no mild keyword appears, the score is 0.

[0127] 1424) The score corresponding to the uncertain trigger word is based on the above scores. ;

[0128] 1425) Output 8-dimensional attribute vector ;

[0129] Example: Sentence: "Multinucleated epithelial cells and a significantly increased nucleocytoplasmic ratio were observed, with coarse granular chromatin; suspected deep nuclear staining was observed."

[0130] Multiple nuclei, which corresponds to a strong positive result and a score of 0.9; High N / C ratio = 0.88, corresponding to a score of 0.95; Coarse-grained, coarse-grained corresponds to moderate, with a score of 0.7; Deep nuclear staining: a questionable basis for correspondences The remaining scores are 0.0. .

[0131] The other steps and parameters are the same as those in one of the specific implementation methods one to five.

[0132] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that: 2) Based on few-shot QA (extraction for VLM / LLM) templates from text Key pathological attributes are obtained, forming attribute vectors. The specific process is as follows:

[0133] A. Build a few-shot example (place it before the command as an example);

[0134] Example 1 (Positive, Strong): Input description (Input description is text) ): It can be seen that the nuclear-cytoplasmic ratio of squamous epithelial cells is significantly increased in many places, the nuclei are clearly deeply stained, and the chromatin is coarsely granulated;

[0135] Expected output:

[0136] {"koilocytosis":0.00,"coarse_chromatin":0.80,"high_nc_ratio":0.90,"hyperchromasia":0.85,"nuclear_irregularity":0.00,"perinuclear_halo":0.00,"multinucleation":0.00,"inflammation":0.00}

[0137] Koilocytosis indicates atypical nuclei, Coarse chromatin indicates coarse chromatin granules, High N / Cratio indicates a high nucleocytoplasmic ratio, Hyperchromasia indicates deep nuclear staining, Nuclear irregularity indicates irregular nuclei, Perinuclear halo indicates a perinuclear halo, Multinucleation indicates multiple nuclei, and Inflammation indicates an inflammatory background.

[0138] Example 2 (Negative / Mild): Input description (Input description is text) Cell morphology was generally regular with no obvious abnormalities; some areas showed mild inflammatory background.

[0139] Expected output:

[0140] {"koilocytosis":0.00,"coarse_chromatin":0.00,"high_nc_ratio":0.00,"hyperchromasia":0.00,"nuclear_irregularity":0.00,"perinuclear_halo":0.00,"multinucleation":0.00,"inflammation":0.40}

[0141] Example 3 (Low-level features): Input description (input description is text) The presence of koilocytes and perinuclear halos suggests HPV-related changes; other features are atypical.

[0142] Expected output:

[0143] {"koilocytosis":0.70,"coarse_chromatin":0.10,"high_nc_ratio":0.10,"hyperchromasia":0.10,"nuclear_irregularity":0.10,"perinuclear_halo":0.75,"multinucleation":0.00,"inflammation":0.10}

[0144] Example 4 (Uncertain / Suspicious): Input description (Input description is text) ): Suspicious nuclear irregularity and suspicious nuclear deep staining, but the evidence is insufficient, and it is recommended to combine more fields of view for evaluation;

[0145] Expected output:

[0146] {"koilocytosis":0.00,"coarse_chromatin":0.10,"high_nc_ratio":0.10,"hyperchromasia":0.40,"nuclear_irregularity":0.40,"perinuclear_halo":0.00,"multinucleation":0.00,"inflammation":0.00}

[0147] Example 5 (Multinucleus + Inflammation): Input description (Input description is text) ): Multinucleated cells and a small amount of coarse granular chromatin are visible against an inflammatory background;

[0148] Expected output:

[0149] {"koilocytosis":0.00,"coarse_chromatin":0.50,"high_nc_ratio":0.00,"hyperchromasia":0.10,"nuclear_irregularity":0.10,"perinuclear_halo":0.00,"multinucleation":0.80,"inflammation":0.70}

[0150] B. Construct the instruction template (single round, constrained output); the specific process is as follows:

[0151] Instructions (for the large language model):

[0152] You will receive a cervical cytology description. Please return only the scores for the 8 attributes specified in the JSON schema, in the range [0,1], rounded to two decimal places; enter 0 for unknown or unmentioned attributes; do not add any explanatory text.

[0153] User content (variables) (Description of this patch):

[0154] describe:

[0155] Please output JSON, where the keys must be in the following order: {"koilocytosis":x1,"coarse_chromatin":x2,"high_nc_ratio":x3,"hyperchromasia":x4,"nuclear_irregularity":x5,"perinuclear_halo":x6,"multinucleation":x7,"inflammation":x8}

[0156] The output should only contain this one line of JSON;

[0157] C. Establish scoring rules (strong / medium / weak → score); the specific process is as follows:

[0158] Strong positive keywords: significant | obvious | high | typical | large amount | severe; strong positive keywords correspond to a score of 0.9; Moderate keywords: moderate | relatively high | multiple | numerous; moderate keywords correspond to a score of 0.7; Mild keywords: mild | few | focal | occasional; mild keywords correspond to a score of 0.5; Only mentioning words without intensity; only mentioning words without intensity corresponds to a score of 0.6; uncertain trigger words correspond to a score based on the above scores. The score for the negation trigger word is 0.0; High N / C Ratio scoring is applied. Numerical hint: When At that time, the high nucleo-cytoplasmic ratio corresponds to a fraction of 0.95; when At that time, the high nucleo-cytoplasmic ratio corresponds to a fraction of 0.85; when At that time, the high nucleo-cytoplasmic ratio corresponds to a fraction of 0.75;

[0159] Note: The numerical labeling of few-shots is consistent with the strong / medium / weak scoring of the rule-based method, which facilitates subsequent calibration and consistency checks of the two results.

[0160] D. Input A, B, and C into the large language model; the large language model outputs an attribute vector. ;

[0161] Output attribute vectors for large language models Perform JSON parsing; if JSON parsing is successful, execute E; if JSON parsing fails, re-execute A, B, C, and D until JSON parsing is successful, then execute E.

[0162] E. Attribute Vectors Each value is cropped to [0,1] (1.2 is cropped to 1, -0.2 to 0, 0.8 remains 0.8, the closest value is 0 and the closest value is 1 and the closest value is 1), and two decimal places are retained to obtain the attribute vector formed by the few-shot QA sampling. .

[0163] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0164] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods One through Seven in that: the context-aware multi-turn question-answering module is used based on fused context representation. The optimal instruction, Prompt, yields the final answer; this module enables the system to engage in fluent and memorable dialogue with the user, which is crucial for achieving "human-machine collaborative" diagnosis; the specific process is as follows:

[0165] S41. Obtaining fusion context representation based on attention mechanism The specific process is as follows:

[0166] S411. After each round of dialogue, the question and answer pair... Encoded into a fixed-dimensional history context vector by a text encoder (such as BERT); represented as:

[0167]

[0168] in, Indicating the first in history Question and answer pairs in turn-based dialogue Encoded context vector; Indicates a text encoder; Indicates the historical number The question in the round-robin dialogue; Indicates the historical number The response to the round-robin dialogue; This indicates vector concatenation;

[0169] S412, When the user requests the current round New problems At that time, the new problems in the current round will be addressed. Input the text encoder to get a new question. ;

[0170] New problems The set of vectors and historical vectors The input is a cross-attention mechanism, and the output is a historical context representation. ; indicates as:

[0171]

[0172] Among them, the historical vector set ; This indicates the question-and-answer session of the first round of historical dialogue. Encoded context vector; This indicates the question and answer session for the second round of historical dialogue. Encoded context vector; Indicating the first in history Question and answer pairs in turn-based dialogue Encoded context vector;

[0173] Fusion Context Representation It contains historical information that is most helpful to the current problem; it can understand that "this" in "how is this cell nucleus" refers to the cell discussed in the previous round, achieving true context awareness;

[0174] S42. Obtaining the optimal instruction Prompt based on the new problem in S41; the specific process is as follows:

[0175] The current round submitted by the user New problems Input a lightweight intent classifier, which will then process the user's new question. Mapped to predefined intent categories, such as "Global Overview," "Region Details," "Feature Localization," and "Diagnostic Suggestions," etc.; represented as:

[0176]

[0177] in, Indicates will The input is the vector output by the text encoder; Indicates weight; Indicates bias; express Activation function layer The activation function layer outputs probability values, such as 0.1, 0.3, 0.5, and 0.2. This indicates that the category corresponding to the highest probability value is selected, for example, the category corresponding to the highest probability value of 0.5; Indicate category; select category The corresponding prompt is the optimal prompt, which guides the VLM to generate the answer in the most appropriate way; the prompt template for the category is written or saved in advance.

[0178] S43, Historical Context Representation Based on S41 and the current round The image patch context generates the final answer. S43's Context outputs the answer. It is not only used for the current round's response, but also stored in the dialogue history (for S41 to remember later); it provides text material to S32. It can also serve as a traceable chain of evidence, as shown in S61 / S62; the specific process is as follows:

[0179] S431. Construct the current round of generation based on the Retrieval-Augmented Generation (RAG) mechanism. Image patch context; the specific process is as follows:

[0180] Search enhancement generation mechanism:

[0181] (1) When a user's question involves uncertain or rare cases, based on the currently viewed image patch The context retrieves the most similar cases from a validated, annotated historical case database. One reference case (image + diagnostic report); The specific process is as follows:

[0182] Calculate the currently interested image patch in the Context Similarity to image patches in historical case database ;

[0183] Calculate the similarity between the optimal instruction Prompt obtained in S42 of the Context and the instruction Prompts in the historical case database. ;

[0184] Calculate the similarity between each augmentation information in the Context (known augmentation information includes: category, source, triggering condition, injection method, and example (covering NILM, ASC-US, LSIL, ASC-H, HSIL)) and augmentation information in the historical case database. ;

[0185] Similarity Similarity Similarity The weighted average is then used to obtain the final similarity score.

[0186] Select cases with high final similarity from the historical case database. One reference case;

[0187] (2) The prompts in the retrieved reference cases are injected as additional context into the prompts in the Context to provide decision-making basis for VLM, thereby significantly improving the accuracy and reliability of the answer and showing the reference basis to the user;

[0188] (3) Based on the Prompt instruction injected in (2), the currently focused image block (or global view), and optional enhanced information to construct the current wheel Image patch context;

[0189] The optional enhancement information includes: category, source, triggering condition, injection method, and example (covering NILM, ASC-US, LSIL, ASC-H, HSIL);

[0190] Category 1: Case Retrievals;

[0191] Category 1 corresponds to the source: annotated cervical pathology case database (images + reports); the trigger condition for Category 1 is: user questions involving rare or uncertain features; the injection method for Category 1 is: concatenating the retrieved case report text into the Prompt; if VLM supports multimodal processing, the images are also input; the retrieved reference cases for Category 1 are: Case A: squamous epithelial cells with regular morphology and uniform staining: diagnosis: NILM; Case B: a small number of atypical cells are seen against an inflammatory background, but not enough to diagnose LSIL, diagnosis: ASC-US; Case C: atypical nuclei and vacuolated cells are seen, consistent with mild lesion characteristics, diagnosis: LSIL; Case D: significantly elevated nucleocytoplasmic ratio, but some features are not enough to completely confirm HSIL, diagnosis: ASC-H; Case E: nucleocytoplasmic ratio >0.8, deeply stained nuclei, coarse chromatin, diagnosis: HSIL;

[0192] Category 2: Knowledge Base Entries;

[0193] Category 2 sources include: WHO classification, Bethesda system, pathology textbooks or guidelines; Bethesda system is the authoritative guideline for cervical cancer diagnosis; the triggering condition for Category 2 is: user questions involving terminology explanations and diagnostic criteria; the injection method for Category 2 is: injecting standard entries into the Prompt in the form of brief definitions; the standard entries are the definitions of diagnostic categories in the authoritative Bethesda system (TBS); the brief definition form is to compress the standard definition into 1-2 sentences, retaining the key diagnostic points; the knowledge base references for Category 2 are: NILM: No intraepithelial lesion or malignant lesion seen; ASC-US: Mild squamous cell abnormality, significance unclear; LSIL: Low-grade squamous intraepithelial lesion, often associated with HPV infection; ASC-H: Atypical squamous cells, HSIL cannot be ruled out; HSIL: High-grade squamous intraepithelial lesion, increased nuclear-cytoplasmic ratio, coarse or uneven chromatin, irregular nuclear membrane.

[0194] Category 3: Statistical Priors;

[0195] Category 3 corresponds to the following sources: historical case data; the triggering conditions for Category 3 are: user inquiries about probability and epidemiological issues; the injection method for Category 3 is: supplementing statistical data or incidence rates in the Prompt; the historical statistical references for Category 3 are: NILM: accounting for 85–90% of cervical smears (≥85, ≤90); ASC-US: accounting for 5–7% of cervical smears (≥5, ≤7); LSIL: accounting for 2–4% of cervical smears (≥2, ≤4); ASC-H: accounting for 0.5–1% of cervical smears (≥0.5, ≤1); HSIL: accounting for 1–2% of cervical smears (≥1, ≤2), requiring close monitoring;

[0196] Category 4: Cross-Modal Cues;

[0197] Category 4 corresponds to the source: ROI detection or attention heatmap; the trigger condition for Category 4 is: user request for localization or visualization support; the injection method for Category 4 is: injecting ROI coordinates, heatmap description (interpretive text of the heatmap based on the VLM model) and additional text into the Prompt; if the model supports it, labeled images can be input simultaneously; the model localization hint for Category 4 is: 3 abnormal cells were detected within the image patch: Cell 1: normal nucleocytoplasmic ratio, regular morphology, corresponding to NILM; Cell 2: slightly irregular, inflammatory background, corresponding to ASC-US; Cell 3: increased nucleocytoplasmic ratio, coarse chromatin, corresponding to HSIL;

[0198] Category 5: Historical Comparisons;

[0199] Category 5 corresponds to the following sources: multiple time-series slides (follow-up) from the same patient; the trigger condition for Category 5 is: user questions involving disease progression; the injection method for Category 5 is: injecting diagnostic comparison information from different time points into the Prompt; the patient follow-up information for Category 5 is as follows: March 2022: diagnosed with NILM, NILM indicates normal cell morphology; June 2023: diagnosed with ASC-US, ASC-US indicates the appearance of a small number of atypical cells; January 2024: diagnosed with LSIL, LSIL indicates mild atypia; June 2025: diagnosed with HSIL, HSIL indicates an increased nucleus-to-cytoplasm ratio and coarse chromatin;

[0200] S432, Represent the historical context of S41 The current wheel constructed with S431 The fine-tuned Visual Large Language Model (VLM) is obtained by inputting the image patch context (Context) into S321. The fine-tuned VLM then generates the final answer. ;

[0201] .

[0202] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.

[0203] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One through Eight in that: the final diagnostic suggestion generation module is used to calculate the average anomaly score based on the attribute vector, perform knowledge graph reasoning based on the average anomaly score, perform probabilistic model prediction based on the attribute vector, and obtain the final diagnostic suggestion based on knowledge graph reasoning and probabilistic model prediction;

[0204] The specific process is as follows:

[0205] This stage will shift from micro-level patch analysis back to the macro-level WSI, leading to the final diagnostic conclusion.

[0206] S51, Attribute vectors based on all valid image patches Calculate the average outlier score for the entire WSI; expressed as:

[0207]

[0208] in, Indicates the first The average anomaly score of the entire WSI across all dimensions; Indicates the first Image blocks Weights, such as cell density based on image patches or model confidence; Represents attribute vector The dimension;

[0209] S52, Attribute vector based on each image patch Calculate the attribute score for each image patch, and generate an interactive diagnostic heatmap covering the entire WSI based on the attribute score of each image patch;

[0210]

[0211] in, Indicates the first Attribute scores for each image patch; Indicates the first Weights of each dimension;

[0212] The darker the color in the image, the higher the risk of lesions in that area, thus guiding doctors to quickly focus on the most suspicious areas.

[0213] S53. Knowledge graph reasoning based on average anomaly scores; the specific process is as follows:

[0214] A knowledge graph was constructed based on authoritative guidelines for cervical cancer diagnosis (such as the Bethesda system). ;

[0215] The average outlier score of the entire WSI As a query, in the knowledge graph The above reasoning is used to generate a clinically logical and interpretable diagnostic path; for example: (high nucleoplasm ratio AND deep nuclear staining AND coarse chromatin) —> implies —> HSIL_Risk_High; —> indicates reasoning, and implies indicates interpretability;

[0216] S54. Probabilistic model prediction based on attribute vectors; the specific process is as follows:

[0217] Based on attribute vectors Constructing feature vectors ;

[0218] eigenvectors Input a pre-trained model into a small, independent probabilistic prediction model (such as a gradient boosting tree or a well-calibrated neural network), and the pre-trained probabilistic prediction model outputs the posterior probability of different lesion grades (NILM, ASC-US, LSIL, HSIL, etc.).

[0219] The pre-trained probabilistic prediction model is trained based on a large amount of historical data;

[0220] S55. A final diagnostic recommendation is obtained based on knowledge graph reasoning and probability model prediction; the specific process is as follows:

[0221] The final diagnostic recommendation combines the results of knowledge graph reasoning and probabilistic model prediction, for example, giving a diagnosis of HSIL (high risk) with a probability of 85%; the main basis is... Significantly elevated nucleocytoplasmic ratio and coarse chromatin features were detected in the image patches. ;

[0222] The final diagnostic recommendations will combine the two reasoning models to achieve a balance between interpretability and accuracy.

[0223] The knowledge graph (based on the Bethesda system) is as follows:

[0224] 1) Construct the nodes and edges of the knowledge graph; specifically:

[0225] 11) The node types of a knowledge graph are:

[0226] Specimen and Quality: Adequate, Inadequate, Obscuring blood / mucus, Inflammation, etc.

[0227] Adequate indicates acceptable (sufficient sample quantity and assessable quality); Inadequate indicates unacceptable / inadequate quality (insufficient cell quantity or severe artifacts); Obscuring blood / mucus indicates blood / mucus obscuring (covering up cytological details);

[0228] Inflammation indicates inflammation (inflammatory background, inflammatory cell infiltration);

[0229] Cytologic Features: High N / C ratio, Hyperchromasia, Coarsechromatin, Nuclear irregularity, Koilocytosis, Multinucleation, Perinuclear halo, Mitosis (atypical), Parakeratosis, Keratinization, Inflammation;

[0230] High N / C ratio indicates a high nucleus-to-cytoplasm ratio; Hyperchromasia indicates deep nuclear staining; Coarse chromatin indicates coarse chromatin granules; Nuclear irregularity indicates irregular nuclear shape (irregular nuclear membrane, abnormal nuclear shape); Koilocytosis indicates koilocytes or atypical nuclei (perinuclear halo); Multinucleation indicates multiple nuclei (one cell containing multiple nuclei); Perinuclear halo indicates perinuclear halo; Mitosis (atypical) indicates atypical mitosis; Parakeratosis indicates parakeratosis; Keratinization indicates keratinization; Inflammation indicates inflammation.

[0231] Composite Patterns: Low-grade pattern, High-grade pattern, Reactive change, etc.

[0232] Low-grade pattern (LGP) represents a low-grade pattern (a typical combination of LSIL features); High-grade pattern (HGP) represents a high-grade pattern (a typical combination of HSIL features); Reactive change represents a reactive change (non-neoplastic changes caused by inflammation or repair);

[0233] Bethesda diagnosis conclusions (Diagnosis Nodes): NILM, ASC-US, LSIL, ASC-H, HSIL, SCC (which can be expanded to AGC, AIS, Adenocarcinoma, etc.);

[0234] NILM (Negative for Intraepithelial Lesion or Malignancy) indicates negative (no intraepithelial lesion or malignancy); ASC-US (Atypical Squamous Cells of Undetermined Significance) indicates atypical squamous cells of undetermined significance; LSIL (Low-grade Squamous Intraepithelial Lesion) indicates low-grade squamous intraepithelial lesion; ASC-H (Atypical Squamous Cells – cannot exclude HSIL) indicates atypical squamous cells, but HSIL cannot be excluded; HSIL (High-grade Squamous Intraepithelial Lesion) indicates high-grade squamous intraepithelial lesion; SCC (Squamous Cell Carcinoma) indicates squamous cell carcinoma; AGC (Atypical Glandular Cells) indicates atypical glandular cells; AIS (Adenocarcinoma in situ) indicates adenocarcinoma in situ; Adenocarcinoma indicates adenocarcinoma.

[0235] Management Recommendations: HPV typing, Colposcopy, Short-interval follow-up, Expert review, etc.

[0236] HPV typing refers to HPV typing test; Colposcopy refers to colposcopy; Short-interval follow-up refers to short-term follow-up (usually 6–12 months); Expert review refers to expert review.

[0237] 12) The edge types of the knowledge graph are:

[0238] Supporting or promoting: Feature → Diagnosis; Feature → Diagnosis indicates that a certain case feature supports or promotes a certain diagnosis; for example, perinuclear halo → LSIL;

[0239] Conflicts or Excludes: Pattern Diagnosis; Pattern Diagnosis indicates that a pattern conflicts with or is mutually exclusive with a diagnosis; for example, an inflammatory pattern. HSIL (inflammatory changes should not be directly diagnosed as HSIL);

[0240] Compositional relationships: Feature → Composite Pattern; Feature → Composite Pattern indicates that multiple low-level features are combined to form a pattern; for example, deep kernel staining → HGP;

[0241] Required conditions: Diagnosis → {essential feature / pattern}; Diagnosis → {essential feature / pattern} indicates that a diagnosis must possess certain essential characteristics or patterns; for example, LSIL → {koilocytes, perinuclear halos};

[0242] Modulations: Quality / Inflammation → Adjusts the weight of evidence; Quality / Inflammation → Adjusting the weight of evidence indicates that some background nodes in the atlas are not direct diagnostic evidence, but will affect the weight of evidence; for example, poor image quality → lowers the upper limit of all feature scores;

[0243] Management Recommendations: Diagnosis → Management; Diagnosis → Management indicates that the diagnostic result can directly trigger downstream management recommendations; for example, HSIL → recommend immediate colposcopy and biopsy.

[0244] 2) Construct the average outlier score for the entire WSI. Mapping to nodes; specifically:

[0245] Input: The average outlier score for the entire WSI ;

[0246] Mapping: Mapping each node With the corresponding attribute components Alignment, and by threshold With temperature parameters Perform smooth activation:

[0247]

[0248]

[0249]

[0250] Use the Sigmoid activation function; For node mapping functions;

[0251] For example: High N / C ratio: Hyperchromasia: Coarse chromatin: Koilocytosis: Nuclear irregularity: Inflammation: ;

[0252] 3) Combining multiple features into composite patterns using fuzzy logic:

[0253] High-level mode :

[0254]

[0255] It is a High-grade pattern; High-grade pattern is a high-level pattern.

[0256] in, For a high nucleus-to-cylinder ratio, For deep staining of the nucleus, It consists of crude nuclear chromatin;

[0257] Low-level mode :

[0258]

[0259] It is a low-grade pattern; where, It is an heterogeneous nucleus. Vacuole / hollowed-out cells;

[0260] reactive or inflammatory pattern :

[0261]

[0262] This is a reactive / inflammatory pattern; among which, Inflammation;

[0263] 4) Rule Template: Define necessary / sufficient condition rules for each Bethesda diagnostic conclusion:

[0264] Conditions for NILM: No significant abnormal features, both HGP and LGP are low;

[0265]

[0266] in, This indicates the score for the NILM diagnostic category;

[0267] The criteria for ASC-US are: mild atypicality exists, but not enough to confirm LSIL;

[0268]

[0269] in, This indicates the score for the ASC-US diagnostic category;

[0270] The criteria for LSIL are: low-grade HPV-related lesions, with Koilocytosis and Perinuclear halo as the core.

[0271]

[0272] in, This indicates the score for the LSIL diagnostic category; Indicates the weights of the low-level pattern (LGP). The weight representing kooilocytosis. The weights representing the perinuclear halo;

[0273] ASC-H criteria: Some features of high-grade lesions are present, but not enough to fully confirm HSIL;

[0274]

[0275] in, This indicates the score for the ASC-H diagnostic category; This represents the threshold, the critical point for determining high-grade lesions; This indicates a fuzzy buffer zone added around the critical point between tolerances / buffer zones to ensure that borderline cases are not misclassified and are uniformly classified as ASC-H.

[0276] Criteria for HSIL: prominent high-grade lesion features;

[0277]

[0278] in, This indicates the score for the HSIL diagnostic category; "Reactive change" indicates a change in responsiveness. This represents the weighting coefficient, used to control the contribution of different features to the HSIL score;

[0279] when There was little change in reactivity; It does not affect the HSIL score; when Significantly altered reactivity; HSIL scores were suppressed;

[0280] Squamous cell carcinoma (SCC):

[0281]

[0282] in, This indicates the score for the SCC diagnostic category; Represents the weighting coefficients, the weights of the High-Level Pattern (HGP), used for pattern combinations in SCC;

[0283] 5) Knowledge graph provides edge weight pattern → Diagnosis (0.7–0.95);

[0284] Pattern → Diagnosis indicates that a composite feature pattern conflicts with or is mutually exclusive with a certain diagnosis;

[0285] The initial values ​​for the edge weights given by the knowledge graph are based on expert experience.

[0286] Composite Pattern = A higher-level node composed of multiple features.

[0287] For example: high-level mode

[0288]

[0289] Low-level mode :

[0290]

[0291] Then, Pattern→Diagnosis maps HGP to HSIL and LGP to LSIL;

[0292] Therefore, Pattern itself is a node composed of Feature, and then it builds an edge with Diagnosis;

[0293] 6) Quality control and inflammation modulation:

[0294] If the WSI quality is insufficient (high inadequate or obcuring artifacts, fewer than 8000 squamous epithelial cells on one WSI image, large number of overlapping cells making individual cells unclear, or blurry image view with blemishes or artifacts, or excessive red cells causing squamous epithelial cells to be obscured): the results are unreliable.

[0295] If the inflammation score on the WSI is greater than 0.6, the confidence level for HSIL / ASC-H is down-adjusted, and the confidence level for ASC-US is up-adjusted. 7)

[0297] 71) The score based on the rule template in 4) is fused with the lower limit of the interval for each category in the historical statistical reference corresponding to category 3 in S431 to obtain the final diagnostic probability. ( );

[0298] The historical statistical reference for Category 3 includes the following intervals for each category: NILM: 85–90% of cervical smears; ASC-US: 5–7% of cervical smears; LSIL: 2–4% of cervical smears; ASC-H: 0.5–1% of cervical smears; HSIL: 1–2% of cervical smears, requiring high attention; the lower limit of the NILM range is 85%; the lower limit of the ASC-US range is 5%; the lower limit of the LSIL range is 2%; the lower limit of the ASC-H range is 0.5%; the lower limit of the HSIL range is 1%; expressed as:

[0299]

[0300] in, This represents the lower bound of the interval fusion category prior for each category in the historical statistical reference corresponding to category 3 in S431; This represents the score of rule template 4); This represents a monotonic function (for scaling the score); Indicates the activation function; Indicates proportionality; This represents the average anomaly score for the entire WSI. Belongs to each category The probabilities; each of the 5 categories has 1 probability, and the sum of the 5 probabilities equals 1. The 5 categories are NILM, ASC-US, LSIL, ASC-H, and HSIL.

[0301] 72) The average outlier score of the entire WSI. The input is a supervised learning model (such as GBDT, NN), which learns the class distribution directly from a large amount of case data. The supervised learning model outputs the final diagnosis probability. ;

[0302] This prediction may be closer to the actual data distribution, but it lacks "interpretability".

[0303] 73) Probability of final diagnosis and final diagnosis probability Weighted fusion is performed to obtain the final diagnostic probability. ; indicates as:

[0304]

[0305] in, Indicates the probability of final diagnosis. The weights; Indicates the probability of final diagnosis. The weights;

[0306] 8) Example of an interpretable path:

[0307] Input features:

[0308]

[0309]

[0310]

[0311] in, The average outlier score representing the high nucleus-to-mass ratio across the entire WSI sheet; This represents the average outlier score for the entire WSI kernel deep staining; This represents the average abnormality score of crude nuclear chromatin across the entire WSI sheet;

[0312] Reasoning result:

[0313] High-level mode :

[0314]

[0315] based on Calculate the score for the HSIL diagnostic category. ≈ 0.33;

[0316] After probability fusion with the ML model (B), the final fusion result (C) is approximately 0.85.

[0317] 0.33 = The result of the rule-based reasoning line (an intermediate product before A).

[0318] The 0.33 here is the HSIL rule score derived from HGP=0.72, without softmax / calibration;

[0319] Only by normalizing 0.33 using softmax + Calib can we obtain the true A.

[0320] 0.85 = Final fusion result (C), which is the final value after probabilistic fusion of rule-based reasoning (A, after calibration) and the ML model (B).

[0321] Output explanation: "Due to the detection of high nucleocytoplasmic ratio (0.82), deep nuclear staining (0.74), and coarse chromatin granulation (0.69), the high-level pattern HGP (0.72) was triggered, and the system infers a diagnosis of HSIL, high risk, probability 85%;".

[0322] The other steps and parameters are the same as those in one of the specific implementation methods one to eight.

[0323] Specific Implementation Method 10: This implementation method differs from Specific Implementation Methods 1 to 9 in that the multi-turn dialogue-based intelligent diagnostic interactive system for cervical cell panoramic images also includes an interactive and traceable diagnostic report output module.

[0324] An interactive and traceable diagnostic report output module is used for multimodal information fusion and bidirectional link interface generation, as well as report generation; the specific process is as follows:

[0325] S61. Multimodal information fusion and bidirectional link interface generation; the specific process is as follows:

[0326] The user interface (UI) will seamlessly integrate the following elements:

[0327] (1) Global Heatmap View: The lowest resolution thumbnail in the full-slice image WSI The heatmap generated by S52 is superimposed on it;

[0328] (2) Two-way link: Clicking on any highlighted area on the heatmap will jump to the corresponding high-resolution image. pieceView; simultaneously, the corresponding image in S32 will be displayed on the right side of the user interface. piece Detailed text, corresponding images in S32 piece The attribute vector corresponding to the text and the corresponding image in S4 piece Historical Q&A records (S4);

[0329] Conversely, click the corresponding image in S32. piece Detailed text, corresponding images in S32 piece The attribute vector corresponding to the text or the corresponding image in S4 piece Historical question and answer records (S4) are used to locate the corresponding image. piece Location;

[0330] (3) Question-and-answer history: Complete dialogue records from S1 to S61 are retained, allowing users to review and trace the entire diagnostic analysis process at any time, such as... Figure 5 As shown;

[0331] S62. Generate a report that combines structured and narrative elements; the specific process is as follows:

[0332] A report that combines structured and narrative elements includes:

[0333] Sample information: WSI ID: XXX, staining method: Pap smear;

[0334] Overall slide quality assessment: Good, sufficient cell quantity, and uniform staining (results from S1-S2);

[0335] High-risk area statistics: A total of N highly suspicious areas were found (see heat map for details);

[0336] Key areas represent images: clickable image patches Patch1, Patch2, and Patch3 (results of S52 and S61);

[0337] Core pathological findings: Atypical cells with significantly increased nucleocytoplasmic ratio, irregular nuclear membrane, and coarse chromatin were detected in three image blocks, consistent with the characteristics of high-grade lesions (results of S52, S53, S54, and S55).

[0338] Diagnostic recommendation: High-grade squamous intraepithelial lesion (HSIL) [Confidence level: XX%] (Results of S53, S54, and S55);

[0339] Recommendations for auxiliary diagnosis: It is recommended that the diagnosis be reviewed and confirmed by a senior pathologist. If clinical conditions permit, HPV genotyping can be considered as supporting evidence (S61 results).

[0340]

[0341] The final output is not a static report, but a dynamic, interactive diagnostic workbench, such as... Figure 4 As shown;

[0342] The report includes not only structured data tables for easy integration with information systems (LIS), but also a narrative summary generated by VLM integration, written in a style similar to that of a human pathologist; see example table 1 for details. Figure 6 As shown.

[0343] The other steps and parameters are the same as those in any of the specific implementation methods one to nine.

[0344] This invention proposes an intelligent diagnostic method for cervical cell panoramic images based on multi-turn question-and-answer interaction. This method deeply leverages the breakthrough capabilities of large-scale vision-language models (VLMs) in multimodal understanding, context awareness, and complex reasoning. It organically integrates multi-scale patch-level feature modeling, fine-grained description generation, dynamic contextual memory mechanisms, and interpretable reasoning-based question answering. The aim is to construct a next-generation cervical cancer auxiliary diagnostic system with high accuracy, strong interactivity, and high reliability. Its overall architecture is as follows: Figure 1 As shown, this system can not only automatically analyze whole-slice scan images (WSI), but also act as a "smart partner" for pathologists, conducting exploratory diagnoses through natural language dialogue.

[0345] The technical solution proposed in this invention, through a series of in-depth innovations from data processing and model building to interaction design, is expected to produce the following significant technical effects, propelling cervical cytology diagnosis into a new intelligent stage:

[0346] This solution represents a paradigm shift from "static reporting" to "dynamic dialogue." Traditional AI-assisted diagnostic systems typically output static heatmaps or fixed classification results, lacking interactivity. This solution, thanks to its multi-turn question-and-answer module, contextual memory mechanism, and dynamic prompt word scheduler, transforms the diagnostic process into an exploratory dialogue between humans and machines. Doctors can ask questions, follow up, and explore any area as if communicating with a colleague. The bidirectional interactive interface further embodies this dynamism, completely replacing the traditional one-way, static image output mode and constructing an unprecedented immersive, dynamic, and interactive new paradigm for AI-assisted diagnosis.

[0347] Achieving "zero-shot" lesion recognition based on understanding rather than segmentation: Traditional computational pathology heavily relies on massive amounts of pixel-level annotations (segmentation masks) to train models, resulting in high costs and lengthy processes. The core innovation of this approach lies in joint image-text modeling. Leveraging the powerful visual understanding and language generation capabilities of Visual Learning Models (VLM), it allows the model to directly "understand" patch images and generate fine-grained pathological descriptions without requiring pre-processed cell segmentation. This means the system can efficiently identify and locate lesion regions through natural language descriptions, significantly reducing reliance on labeled data and demonstrating strong "zero-shot" or "few-shot" learning potential.

[0348] One of the bottlenecks in AI applications in the medical field is the lack of trust caused by the "black box" problem. This solution aims to improve the transparency of the system and doctors' trust through multiple mechanisms. First, the generated natural language descriptions present the model's judgment process in a human-understandable way. Second, the introduced knowledge graph reasoning ensures that diagnostic recommendations follow accepted clinical logic. Most importantly, RAG retrieval-enhanced generation can retrieve similar cases from historical databases as "precedents" when the model encounters uncertainty, providing doctors with a solid and traceable chain of diagnostic evidence, thereby significantly improving doctors' understanding and trust in the AI ​​system's output.

[0349] Empowering comprehensive image analysis, the system significantly improves screening efficiency and diagnostic accuracy. Through fully automated preprocessing and efficient screening processes, the system can quickly eliminate massive amounts of irrelevant information. More importantly, the generated panoramic diagnostic heatmap can guide doctors' attention to the most suspicious high-risk areas in seconds, avoiding the huge time consumption of manually reviewing each field of view. Combined with the precise and quantitative assessment of pathological features by VLM, the system can not only significantly reduce image review time (estimated to be reduced by 50%-70%), but also effectively reduce missed diagnoses and misdiagnoses caused by visual fatigue or experience differences, thereby improving the overall efficiency and accuracy of large-scale cervical cancer screening.

[0350] This solution establishes a general pathological diagnostic framework with domain-adaptive capabilities, exhibiting high versatility and transferability. The "WSI preprocessing + Patch-level VLM description generation + multi-round interaction + diagnostic aggregation" approach is not limited to cervical cytology. Leveraging efficient LoRA fine-tuning technology, by simply replacing the training data and knowledge graph / rule base for a specific domain, the system can quickly adapt and be extended to other types of pathological image-assisted diagnostic scenarios, such as HER2 interpretation in breast cancer, histological grading of liver cancer, and subtype classification of lung cancer, demonstrating significant application extension value.

[0351] The modular and easily integrated clinical deployment solution system is designed with a modular approach, featuring clear technical logic, low coupling, and easy independent upgrades and maintenance. In particular, the defined structured report generation function outputs data in a format that seamlessly integrates with existing hospital information systems (HIS), laboratory information systems (LIS), or pathology information systems (PIS). This high degree of scalability and system integration ensures that this invention is not merely a technical prototype, but a mature solution that can be easily and quickly deployed and deliver value in real-world clinical environments.

[0352] Key points and points to be protected in this invention

[0353] The key innovations that distinguish this invention from existing technologies, and the core technologies we seek to protect, are mainly reflected in the following aspects. These technologies collectively constitute a complete, efficient, and reliable new framework for interactive pathological diagnosis.

[0354] Application and Logical Framework of Multi-Turn Text-Image Dialogue Question-Answer Mechanism in Medical Image Diagnosis

[0355] One of the core contents that this invention aims to protect is a complete, closed-loop multi-turn image-text dialogue logic framework specifically designed for medical image diagnosis. It is not simply a matter of grafting question-and-answer functionality onto an image viewer, but rather a collaborative workflow encompassing intent recognition, context fusion, image localization, and response generation.

[0356] Detailed description of protection points: We are seeking to protect a complete logical framework as follows:

[0357] Intent-Driven Orchestration: When a user asks a question, the system does not immediately generate an answer. Instead, it first categorizes the question intent using a dynamic prompt word scheduler (e.g., "global overview", "feature localization", "detailed description").

[0358] Context-Dynamic Fusion: Based on the identified intent, the system extracts and fuses the most relevant historical context vectors from the dialogue history memory cache in parallel using an attention mechanism to correctly resolve referential relationships (such as "How is it?").

[0359] Visual Evidence Association: Based on the question content and context, the system automatically locates the most relevant single or multiple image patches on WSI and retrieves their visual information as the core basis for the answer.

[0360] Contextualized Answer Generation: Finally, the Visual Language Model (VLM) integrates image content, the current question, historical context, and dynamically generated prompts to generate an accurate textual or multimodal answer that is highly relevant to the current dialogue context.

[0361] This complete and structured information processing chain, from "asking a question" to "answering a question," is the core innovation of this invention in terms of interaction method, and we advocate the protection of its overall logical framework.

[0362] Natural language description and automatic, bidirectional image patch association and response generation technology;

[0363] The second key point that this invention aims to protect is a technology for achieving automatic, persistent, and bidirectional binding between microscopic regions (patches) of pathological images and their semantic natural language descriptions.

[0364] Detailed description of the protection points: We seek to protect the technical methods implemented in S3:

[0365] Automatic generation and binding from visual to text: For each valid image patch, the system automatically generates a fine-grained pathological description and a quantified attribute vector through a domain-fine-tuned Visual Model (VLM). The system creates and solidifies this information within its internal data structure. and A strong correlation mapping relationship between them.

[0366] Real-time response from text / interaction to visual: Based on the above binding relationship, the system realizes bidirectional query response.

[0367] Spatial location response: When a user clicks on an area on the interface, the system can instantly retrieve and display the description associated with it. and attribute vector ;

[0368] Semantic localization response: When a user asks, "Which cells in the image have deeply stained nuclei?", the system can retrieve all... The text content is back-located to all image patches containing the description "kernel deep staining" and highlighted on WSI.

[0369] This technology, which automatically creates "identity cards" (i.e., natural language descriptions) for massive images and uses them as a bridge to achieve seamless, bidirectional retrieval and interaction between vision and semantics, is the foundation for achieving efficient and intuitive diagnosis and constitutes a core protection point.

[0370] Dialogue memory caching in structured embedding and attention-weighted methods for sequential question answering

[0371] The third key point that this invention seeks to protect is a structured embedding and dynamic weighting method of dialogue history based on an attention mechanism, specifically designed to maintain the coherence of medical diagnostic dialogues.

[0372] Detailed Explanation of the Protection Points: We seek to protect specific, defined memory processing mechanisms, rather than the general concept of "memory." These mechanisms include:

[0373] Structured embedding: embedding each round of question-and-answer pairs The text encoder uniformly encodes the data into a structured, fixed-dimensional context vector. And store them in an ordered memory bank.

[0374] Dynamic weighted attention: When dealing with new problems At that time, the system does not simply splice or average historical information. Instead, it encodes the current problem. As a query, the entire memory is used as a key and value, and computation is performed through a cross-attention module.

[0375] Context vector generation: The output of the attention mechanism is a weighted fused context representation. In this vector, the historical dialogue fragments most relevant to the current question are given the highest weight, while the weights of irrelevant information are suppressed.

[0376] This method, which uses attention mechanisms to dynamically weight the structured dialogue history to generate a highly focused contextual representation most relevant to the current issue, is a key technology for solving complex referential and logical coherence problems in medical dialogues. We advocate for the protection of the method itself.

[0377] Diagnostic output format for unlabeled identification of abnormal regions and explanatory text synthesis.

[0378] The fourth key point that this invention aims to protect is a novel form of diagnostic information output that combines abnormal region identification technology that does not require preset classification labels with real-time synthesized natural language interpretation.

[0379] Detailed Explanation of Protection Points: We seek to protect a combined output format of diagnostic results and its generation method.

[0380] Classification-Free Identification: As described in S3, the system determines whether a region is "abnormal" not by classifying it into a pre-defined label (such as HSIL, LSIL), but by analyzing its visual content through VLM and generating descriptive text. Does the sample contain certain key pathological descriptive terms (such as "nucleoplasmic imbalance" or "coarse chromatin")? This is an open-ended anomaly identification method based on semantic understanding.

[0381] Explanatory Text Synthesis: For regions identified as anomalous, the system presents the user not with a simple label, but with a descriptive text synthesized in real time by the Visual Model based on the specific visual features of that region. This text provides a direct explanation of "why this is abnormal".

[0382] Therefore, what is being protected is this diagnostic unit output format of "visual highlighting (unlabeled) + dynamically generated text explanation." It fundamentally changes the traditional AI's "black box" model of "giving answers without explanations," and is the core of building trustworthy AI.

[0383] A diagnostic report generation method that integrates image summarization, question-answering history, abnormal signs, and reasoning chains

[0384] The fifth key point that this invention seeks to protect is a systematic method that can automatically integrate multi-source, heterogeneous information to generate a comprehensive and traceable final diagnostic report.

[0385] Detailed Explanation of Protection Points: We are not seeking to protect the report itself, but rather the complete method described in S6 that automatically integrates the following four types of information to generate the report: Global Image Summary: Quantitative abnormality scores and diagnostic heatmaps of the entire image WSI are extracted to provide a macroscopic summary of the severity and distribution of lesions.

[0386] Complete interaction history: The system fully references the recorded human-computer question-and-answer history as complete evidence of the traceability of the diagnostic process.

[0387] Key abnormal signs extraction: The system automatically extracts the core descriptive text generated by VLM from the most abnormal patches as “pathological evidence” to support the diagnostic conclusion.

[0388] Explicit reasoning chain: Clearly link the final diagnostic recommendation (such as "HSIL") with its reasoning basis (such as "According to the knowledge graph, 'high nucleocytoplasmic ratio' and 'coarse chromatin' strongly point to HSIL") to show the logical path that leads to the conclusion.

[0389] This systematic method, which organically integrates and automatically arranges macro-quantitative data, complete interactive processes, micro-textual evidence, and top-level logical reasoning to generate a multi-dimensional, self-explanatory, and traceable diagnostic report, is the key integrated innovation of this invention in its top-level design. We advocate protection of its overall generation method.

[0390] Regarding the technical solution, are there any alternative solutions that can achieve the same purpose of the invention?

[0391] To achieve the core objective of this invention—"to construct an interactive and interpretable cervical cell image-assisted diagnostic system"—there are theoretically several alternative technical approaches. However, these alternatives differ fundamentally from the solution proposed in this invention in terms of implementation, effectiveness, and system integration, and often have significant limitations. The following analysis and comparison of several major alternatives will highlight the innovation and superiority of this invention.

[0392] Alternative Solution 1: Pipeline-based Feature Extraction and Text Generation Solution

[0393] This was a more traditional and modular implementation method before the widespread adoption of large-scale visual-language models (VLMs).

[0394] Technical path description

[0395] Step 1: Independent Visual Feature Extraction. The image patch is analyzed using a pure computer vision (CV) model (such as CNN-based ResNet, EfficientNet, or U-Net for segmentation). The output of this model is structured, non-textual data, such as a classification label (e.g., "HSIL"), a set of quantized feature scores, or a pixel-level segmentation mask.

[0396] Step 2: Independent Natural Language Generation. The structured data output from Step 1 is then input into an independent Natural Language Generation (NLG) model (such as GPT or T5) or a template-based system.

[0397] Step 3: Text synthesis. The NLG model "translates" the input structured data into a natural language description based on preset rules or templates.

[0398] Comparison and limitations with the present invention;

[0399] Information is severely lost during transmission: this is the most fatal flaw. The structured data extracted by the CV model is a lossy compression of the original image information. The NLG model cannot see the original image; all the information it relies on comes from the result of this compression. Any visual details in the image that were not predefined and extracted by the CV model will be lost forever and cannot be mentioned or retrieved in subsequent text generation or question-answering stages.

[0400] Interactivity and flexibility are extremely poor: Because the text model cannot see images, users cannot ask open-ended visual questions. For example, a user cannot ask, "What is that darker-colored cluster of cells in the top left corner of the image?" because the NLG model does not know what "top left corner" or "darker-colored" visually corresponds to. Interaction is limited to restating extracted features.

[0401] Conclusion: The end-to-end VLM scheme adopted in this invention allows the language model to directly "see" pixels, achieving a deep fusion of vision and language. This fundamentally overcomes the information loss and interaction barriers of pipelined schemes, representing a paradigm shift.

[0402] Alternative Solution 2: Anomaly Region Detection Scheme Based on Traditional Classification / Segmentation Models

[0403] This solution can locate abnormal areas, but it cannot support true natural language interaction.

[0404] Technical path description;

[0405] Model construction: Use the Multiple Instant Learning (MIL) framework for WSI-level classification, or use models such as U-Net for pixel-level cell segmentation and classification.

[0406] Output results: The model ultimately outputs a WSI-level diagnostic classification, along with a heatmap highlighting the abnormal areas, or directly outlines the identified abnormal cells on the map.

[0407] Comparison and limitations with the present invention;

[0408] Lack of fine-grained interpretability: These models can only tell doctors "which areas are important" or "which cells are abnormal," but cannot explain "why these areas are important" in natural language. What doctors see is still a "black box" of highlighted areas that they need to interpret themselves.

[0409] The interaction is extremely limited: the system's entire function is classification and location, lacking any dialogue capabilities. Doctors cannot ask questions or explore the highlighted areas in any way, severely restricting the depth and breadth of human-computer interaction.

[0410] Conclusion: Traditional segmentation or classification models can assist in the localization of abnormal regions, but they cannot support interactive question-and-answer based on natural language, which is far from the goal of this invention to build a "human-machine collaborative intelligent partner". This invention, by introducing VLM for fine-grained description, fundamentally shifts the core function of the system from "classification" to "understanding and dialogue".

[0411] Alternative Option 3: A description scheme based on a CBIR (Content-Based Image Retrieval) system;

[0412] This approach provides information by finding similar cases, but lacks the ability to dynamically analyze current cases.

[0413] Technical path description:

[0414] Database construction: A large, annotated database of pathological images is pre-built, with each (or each) patch having a standard pathological description written by an expert.

[0415] Retrieval and Matching: When a user queries a new image patch, the system uses Content-Based Image Retrieval (CBIR) technology to find the N most visually similar historical cases from the database.

[0416] Results presentation: The system displays pre-written static descriptions of these similar cases to the user for reference.

[0417] Comparison and limitations with the present invention:

[0418] Lack of dynamic response and realistic analysis capabilities: This approach analyzes similar past images rather than current ones. It cannot describe or respond to unique, subtle, or rare features in the current patch; its response is static and indirect.

[0419] Excessive database dependency: The system's effectiveness depends entirely on the size, quality, and coverage of the backend database. The system will be unable to process lesion types or feature combinations not included in the database.

[0420] Conclusion: Image retrieval systems or static description generation methods can partially accomplish cell description tasks, but they lack dynamic analysis and real-time response capabilities for the current query, failing to meet the requirements of specific and real-time analysis in clinical diagnosis. The VLM scheme of this invention can perform real-time, dynamic "image interpretation" on any input image, possessing fundamental advantages.

[0421] In summary, although there are the above-mentioned and other scattered alternative technologies, they either suffer from irreparable information loss in core processes (pipeline solutions), or their design goals contradict the interactive and interpretive concept of this invention (classification / segmentation models), or they have shortcomings in dynamic analysis in key capabilities (image retrieval solutions).

[0422] The core logic upon which this invention relies—the integrated framework of "multi-turn text-image dialogue + contextual memory + anomaly aggregation and reasoning generation"—cannot be fully covered by existing single-point technologies or simple combinations. The originality and irreplaceability of this invention are reflected in:

[0423] It is not a simple optimization of the above solutions, but a creative integration of end-to-end visual language understanding, attention-based long-range conversational memory, and knowledge-based diagnostic reasoning into a completely new, organic, and collaborative overall framework. This framework achieves a deep unification of the three major goals of "interaction," "interpretation," and "diagnosis," which cannot be achieved by existing alternative technological paths, and therefore has significant creativity and progressiveness.

[0424] What are the advantages of this invention compared to the closest existing technology?

[0425] Compared with existing technologies, this invention is not a simple addition of functions or performance improvement, but a fundamental breakthrough in human-computer interaction paradigms, transparency of the diagnostic process, depth of information processing, and system intelligence. The closest existing technologies can be mainly classified into two categories: one is traditional computer vision (CV) diagnostic systems represented by multiple instance learning (MIL) or segmentation models, whose core is classification and localization; the other is basic vision-language model (VLM) applications, such as static image description generation.

[0426] The advantages of this invention are mainly reflected in the following aspects:

[0427] A fundamental shift in interaction patterns: from "static reporting" to "dynamic exploration";

[0428] Existing Technology: The closest CV diagnostic systems (such as MIL-based classifiers) output static results, typically a heatmap containing diagnostic labels (e.g., "HSIL") and marking suspicious areas. Users (doctors) can only passively accept this result, unable to ask follow-up questions or explore further. Human-computer interaction is unidirectional and fixed. Advantages of this Invention: This invention, through its multi-turn question-and-answer module, transforms human-computer interaction from a unidirectional "report mode" to a bidirectional "dialogue mode." Doctors can freely ask questions, follow up, and explore in depth using natural language, focusing on any visual feature on the WSI, for example: "Please zoom in on the darkest area in the heatmap and describe the morphology of the cell nuclei within it." The system can respond in real time, turning the diagnostic process into a dynamic, doctor-led exploration. This greatly enhances the system's usability and the doctor's sense of participation.

[0429] Transparency and interpretability of the diagnostic process: from "black box" to "white box";

[0430] Existing technology: Traditional computer vision (CV) systems, whether classification or segmentation models, present a decision-making process that is essentially a "black box" to the user. The system can highlight a region, but it cannot explain "why" that the region is considered abnormal in language that humans can understand. Doctors need to rely on their own experience to reinterpret the AI's judgment, resulting in high trust costs.

[0431] Advantages of this invention: This invention achieves "white-box" diagnostics through two main mechanisms: Fine-grained description based on VLM: The system generates not a label for each abnormal region, but a specific, pathologically logical description in natural language (e.g., "This cell cluster shows a significant increase in nucleoplasmic ratio and deep nuclear staining"). This description itself is the most direct and fundamental explanation for the diagnostic judgment.

[0432] Knowledge-based reasoning chain: The final diagnostic recommendation is clearly linked to the pathological features and knowledge graph rules on which it is based.

[0433] These two points together make the "inner workings" of AI completely transparent, so that doctors not only know what it does, but also why it does it, thus fundamentally solving the trust problem in the application of AI in medicine.

[0434] The depth and losslessness of information processing: from "information silos" to "deep integration";

[0435] Existing technologies: Some attempts to combine computer vision (CV) with text in a "pipeline" manner suffer from a fatal flaw: information is severely lost during module transfer. The CV model first compresses the image into structured data (such as classification labels or feature vectors), and the text model then generates a description based on this secondary information. In this process, the rich, undefined visual details of the original image are permanently lost.

[0436] Advantages of this invention: This invention employs an end-to-end Visual-Language Model (VLM) as its core, achieving deep and synchronous processing of visual and linguistic information. Within the unified model space of the VLM, text queries can directly apply to pixel-level visual features. This means there are no intermediate steps that result in information loss; the model can capture and describe any subtle visual features in an image, regardless of whether they are predefined. This lossless information processing method is the foundation for achieving accurate and flexible interaction.

[0437] Context-aware intelligent interaction: from "isolated question and answer" to "coherent dialogue";

[0438] Existing technology: Basic VLM applications are typically question-and-answer based, lacking the ability to remember the dialogue history. If a user asks consecutive questions, such as first asking "What's wrong with this area?" and then "What's its nucleocytoplasmic ratio?", the system may not understand that "it" in the second question refers to the area discussed in the first question. This "goldfish-like memory" cannot support logical and in-depth diagnostic communication.

[0439] Advantages of this invention: This invention endows the system with long-term, efficient contextual understanding capabilities through a dialogue memory caching based on an attention mechanism. The system can structurally encode the dialogue history and dynamically assign the highest weight to the most relevant historical information when answering new questions. This enables the system to accurately understand complex referential relationships, maintain the logical coherence of the dialogue, and achieve truly intelligent, multi-turn dialogue.

[0440] Completeness and traceability of diagnostic reports: from "single conclusion" to "process panorama";

[0441] Existing technology: The output diagnostic reports usually only contain final conclusive information, such as a diagnostic label and a static heatmap, completely missing the intermediate analysis process and interactive exploration history that led to the conclusion.

[0442] Advantages of this invention: The report generation method defined in this invention is original. It can automatically integrate and generate a comprehensive report that includes a summary of macroscopic images, textual evidence of key abnormal signs, an explicit chain of diagnostic reasoning, and a complete history of doctor-patient interactions. This report is not only a conclusion, but also a panoramic and traceable record of the entire diagnostic process. This has immeasurable value for medical quality control, teaching and training, and doctor-patient communication.

[0443] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A multi-turn dialogue-based intelligent diagnostic interactive system for cervical cell panoramic imaging, characterized by: The system includes: The module includes: image patch acquisition module, image segmentation and effective region filtering module, image-text pair construction and pathological attribute acquisition module, context-aware multi-turn question answering module, and final diagnosis suggestion generation module. The image block acquisition module is used to acquire the whole slice image WSI, and to preprocess the acquired whole slice image WSI to obtain the preprocessed whole slice image WSI. The image segmentation and effective region filtering module is used to divide the preprocessed whole slice image WSI into blocks to obtain image blocks, and filter the image blocks to obtain effective image blocks. The image-text pair construction and pathological attribute acquisition module is used to construct instructions (Prompt) for the obtained valid image patches. The constructed instructions (Prompt) and valid image patches are input into the Visual Large Language Model (VLM), and the Visual Large Language Model (VLM) outputs text. Pathological attributes are parsed and quantified from the text to form attribute vectors. Context-aware multi-turn question answering modules are used for fusion context representation And the optimal instruction Prompt yields the final answer; The final diagnostic suggestion generation module is used to calculate the average anomaly score based on the attribute vector, perform knowledge graph reasoning based on the average anomaly score, perform probabilistic model prediction based on the attribute vector, and obtain the final diagnostic suggestion based on knowledge graph reasoning and probabilistic model prediction.

2. The multi-turn dialogue-based intelligent diagnostic interactive system for cervical cell panoramic imaging according to claim 1, characterized in that: The image block acquisition module is used to acquire the whole slice image WSI, and preprocess the acquired whole slice image WSI to obtain the preprocessed whole slice image WSI; the specific process is as follows: S11. Use the OpenSlide open-source library to obtain the full slice image WSI and its metadata; The metadata of a full slice image (WSI) includes: the physical size of the WSI, the scan magnification, and the resolution of each level in the image pyramid; The full-slice image WSI is explicitly constructed as an image pyramid structure: in, WSI for full slice images; This is the highest resolution image in WSI for a full-slice image; This is the second highest resolution image in WSI for a full-slice image; This is the lowest resolution thumbnail in the WSI full-slice image; S12. Preprocess the acquired full-slice image WSI to obtain the preprocessed full-slice image WSI; the specific process is as follows: S121. Use the standard color normalization algorithm to process the lowest resolution thumbnail in the full-slice image WSI. The correction is performed to obtain the corrected whole-slice image WSI; S122. Input the corrected whole-slice image WSI into the U-Net segmentation model. The U-Net segmentation model outputs a binarized tissue region mask. ; Binarized tissue region mask This indicates that the corrected full-slice image WSI shows a blank background area; Binarized tissue region mask The corrected whole-slice image WSI indicates the cervical cell region; Preservation of binarized tissue region mask The corrected full-slice image WSI.

3. The multi-turn dialogue-based intelligent diagnostic interactive system for cervical cell panoramic imaging according to claim 2, characterized in that: The image segmentation and effective region filtering module is used to divide the preprocessed whole slice image WSI into blocks to obtain image blocks, and filter the image blocks to obtain effective image blocks. S21. Divide the preprocessed full-slice image WSI from S12 into blocks to obtain image patches; the specific process is as follows: Select the middle layer image in the image pyramid structure. ; An overlapping sliding window strategy is used to separate the intermediate layer images. Cut into fixed-size image blocks: in, To make the intermediate layer image A collection of image blocks cut into fixed sizes; For set The first image patch; For set The second image patch in the middle; For set The Middle Image blocks; For set The Middle Image blocks; To make the intermediate layer image The total number of image blocks cut into fixed sizes; To make the intermediate layer image The first image block cut into fixed size indivual; ; , For real numbers, For the first Image blocks of high, For the first Image blocks The width, 3 is the first Image blocks RGB3 channels; S22. Filter the image blocks to obtain valid image blocks; the specific process is as follows: S221, the set of image patches obtained in S21 Calculate tissue coverage for each image patch ; Set tissue coverage threshold , ; Eliminate Image blocks, preserved Image blocks; The image patch set obtained from S21 Calculate tissue coverage for each image patch The expression is: in, Indicates the first Image blocks The corresponding mask matrix; This represents the pixel threshold, with a value of 0.

5. Represents image blocks The mask corresponding to each pixel is greater than the pixel threshold. The number of image blocks; Indicates the first Image blocks Organizational coverage; S222, Reservation of S221 Calculate the information entropy for each image patch; set the information entropy threshold; Image blocks with information entropy less than the information entropy threshold are discarded as valid image blocks.

4. The multi-turn dialogue-based intelligent diagnostic interactive system for cervical cell panoramic imaging according to claim 3, characterized in that: The image-text pair construction and pathological attribute acquisition module is used to construct a Prompt instruction for the obtained valid image patches. The constructed Prompt instruction and the valid image patches are input into the Visual Large Language Model (VLM), and the VLM outputs text. Pathological attributes are parsed and quantified from the text to form attribute vectors. The specific process is as follows: S31. Construct the instruction Prompt for each valid image patch obtained in S22; The Prompt command is: Basic Description: Please describe the first... The number of valid image patches includes: First, evaluating the first... The quality and cell density of each effective image patch; then, the quality and cell density of the first effective image patch are identified and described. Cell types of the first valid image patch; finally, please pay attention to the first... Does the valid image block contain abnormal cell features? Feature query: Do each valid image patch have cells with the following characteristics: 1) enlarged and irregular nuclei; 2) increased nucleus-to-cytoplasm ratio; 3) granular or unevenly distributed chromatin; 4) presence of deep nuclear staining? S32, the first obtained in S22 The Prompt instruction for the first valid image block and the... Each valid image patch is input into the Visual Large Language Model (VLM), which outputs text. Pathological attributes are parsed and quantified from the text to form attribute vectors. The specific process is as follows: S321. Use LoRA technology to fine-tune the Visual Large Language Model (VLM) to obtain the fine-tuned Visual Large Language Model (VLM). The specific process is as follows: S3211, The forward propagation process of a certain layer of the Visual Large Language Model (VLM) becomes: in, It is a low-rank matrix. ; It is a low-rank matrix. ; The token vector is the input to a certain layer of the visual large language model; It is a scaling constant; This is the weight matrix of the visual large language model; For rank; The dimension of the token vector input to a certain layer; The dimension of the token vector output by a certain layer; This is the output vector after incorporating the low-rank LoRA update. S3212. Constructing a multi-task joint loss function for a visual large language model (VLM). Multi-task joint loss function Convergence yields the fine-tuned Visual Large Language Model (VLM). Multi-task joint loss function for: in, This represents the text generation loss, which is the cross-entropy loss. express The weight hyperparameters; Indicates classification loss; express The weight hyperparameters; This represents the regression loss, which is the mean squared error loss. express The weight hyperparameters; S322, the first obtained in S22 The Prompt instruction for the first valid image block and the... The fine-tuned Visual Large Language Model (VLM) takes n valid image patches as input and outputs the nth valid image patch. Each image and text pair corresponds to a text. ; in, This represents the fine-tuned Visual Large Language Model (VLM). It represents a coherent natural language description; S323. Parse and quantify pathological attributes from the text to form attribute vectors; the specific process is as follows: From the text The pathological attributes are obtained, forming an attribute vector. ; indicates as: in, This represents the probability of Koilocytosis (a type of nuclear malformation). ; The Coarse Chromatin score indicates the coarse granulation of chromatin. ; This indicates a high N / C ratio score. ; This indicates the Hyperchromasia score, which represents nuclear hyperchromatography. ; Indicates the nuclear irregularity score. ; This represents the probability of a vacuolated / hollowed-out cell (Perinuclear Halo). ; This represents the probability of multinucleation. ; Indicating an Inflammatory Background score. .

5. The multi-turn dialogue-based intelligent diagnostic interactive system for cervical cell panoramic imaging according to claim 4, characterized in that: The text The pathological attributes are obtained, forming an attribute vector. The specific process is as follows: 1) From text based on template matching rules The pathological attributes are obtained, forming an attribute vector. ; 2) Based on few-shot QA templates from text The pathological attributes are obtained, forming an attribute vector. ; 3) The attribute vector formed in step 1) based on template matching rules and 2) the attribute vector formed by sampling few-shot QA By performing rule-based fusion, the final attribute vector is obtained. ; The specific process is as follows: in, Represents the weight vector. .

6. The multi-turn dialogue-based intelligent diagnostic interactive system for cervical cell panoramic imaging according to claim 5, characterized in that: In step 1), the template matching rule is used to obtain text The pathological attributes are obtained, forming an attribute vector. The specific process is as follows: Template matching rules include keywords | regular expression dictionary, negation | uncertain trigger words, and scoring rules; 11) Keywords | Regular expression dictionary; The specific process is as follows: Probability of heteronuclear Koilocytosis The keywords are: heterogeneous nucleus | heterogeneous nucleus | koilocyt | hollowed-out nucleus | perinuclear halo cell; where "|" indicates or; Coarse Chromatin Score The keywords are: coarse chromatin | coarse chromatin | large chromatin | uneven chromatin distribution | chromatin clumps / coarse chromatin | clumped chromatin; where " / " is the separator between Chinese and English keywords; Here, coarse chromatin represents coarse nuclear chromatin; clumped chromatin represents clustered chromatin. High N / C Ratio Score The keywords are: increased nucleocytoplasmic ratio | elevated nucleocytoplasmic ratio | significantly increased nucleocytoplasmic ratio | high nucleocytoplasmic ratio / high N / C; where high N / C indicates a high nucleocytoplasmic ratio; the regular expression dictionary value suggestion is: regular expression capture. ; Hyperchromasia score The keywords are: nuclear hyperchromatography | nuclear staining intensification | nuclear overstaining / hyperchrom; Among them, hyperchrom represents nuclear hyperchromatography; Nuclear Irregularity Score The keywords are: irregular nuclear shape | irregular nuclear outline | irregular nuclear membrane | irregular nuclear notch / irregular nuclear; where irregular nuclear indicates an irregular nuclear shape; Perinuclear Halo probability of vacuolated / hollow cells The keywords are: perinuclear halo | vacuolated cells | vacuolation | halo; where halo represents perinuclear halo; Multinucleation probability The keywords are: multinucleat | multinucleated giant cell | binucleat | trinucleat | multinucleatization; where multinucleat represents multinucleatization; multinucleat represents multinucleatization. Background inflammation (Inflammatory Background Score) The keywords are: inflammatory background | inflammatory cells | neutrophilia | lymphocytosis | purulent background | inflammatory exudate / inflamm; where inflamm represents inflammation; 12) Negation | Indeterminate trigger words; the specific process is as follows: The trigger words for negation are: not seen | no obvious | not found | not supported | does not conform | not considered | excluded; Uncertain trigger words include: suspicious, pending, tendency, possible, consider, mild; 13) Scoring rules; the specific process is as follows: Strongly positive keywords: significant | obvious | highly | typical | large amount | severe; The score corresponding to strongly positive keywords is 0.9; Medium-level keywords: moderate | high | multiple | numerous; The corresponding score for medium-level keywords is 0.7; Mild keywords: mild | few | localized | occasional; Mild keywords correspond to a score of 0.5; Only mentioning words with no intensity; mentioning only words with no intensity corresponds to a score of 0.6; The score corresponding to the uncertain trigger word is based on the above score. ; The score corresponding to the negation trigger word is 0.0; High N / C Ratio Score Numerical hints: when At that time, the high nucleo-cytoplasmic ratio corresponds to a score of 0.95; when At that time, the high nucleo-cytoplasmic ratio corresponds to a fraction of 0.85; when At that time, the high nucleo-cytoplasmic ratio corresponds to a fraction of 0.75; 14) Rule evaluation process; the specific process is as follows: 141) Regarding the text Preprocessing is performed to obtain the preprocessed text. The specific process is as follows: Text Convert each word in the text to half-width characters to obtain the text after half-width conversion. ; Convert the text to half-width characters Standardize the terminology for synonyms in the text; Text after standardization of terminology Consistent use of capitalization in English text; Text with consistent capitalization Denoising is performed to obtain the preprocessed text. ; 142) For the preprocessed text The scoring process is as follows: 1421) For the preprocessed text First, check for negation. If a negation trigger word exists, set the corresponding word to 0.0 and execute step 1422); if no negation trigger word exists, execute step 1422. 1422) Regarding the text Detection of high N / C ratio; If a high N / C ratio exists, the score is based on the high N / C ratio. The numerical prompt will retrieve the corresponding score; execute 1423). If a high N / C ratio does not exist, the high N / C ratio score is... The value is 0.0, so execute 1423); 1423) Assign corresponding scores to the results of 1422) according to the keywords of strong positive / moderate / mild, and then perform 1424); the specific process is as follows: If a strong positive keyword appears, the corresponding score is taken; if multiple strong positive keywords appear simultaneously, the maximum score is taken; if no strong positive keyword appears, the score is 0. If a medium-level keyword appears, the corresponding score is taken; if multiple medium-level keywords appear simultaneously, the maximum score is taken; if no medium-level keyword appears, the score is 0. If a minor keyword appears, the corresponding score is used; if multiple minor keywords appear simultaneously, the maximum score is used; if no minor keywords appear, the score is 0. 1424), the score corresponding to the uncertain trigger word is based on... ; 1425) Output 8-dimensional attribute vector .

7. The multi-turn dialogue-based intelligent diagnostic interactive system for cervical cell panoramic imaging according to claim 6, characterized in that: In step 2), the text is based on the few-shot QA template. The pathological attributes are obtained, forming an attribute vector. The specific process is as follows: A. Create a few-shot example; Example 1: Input description: Multiple squamous epithelial cells showed a significantly increased nuclear-cytoplasmic ratio, with markedly darkly stained nuclei and coarsely granulated chromatin; Expected output: {"koilocytosis":0.00,"coarse_chromatin":0.80,"high_nc_ratio":0.90,"hyperchromasia":0.85,"nuclear_irregularity":0.00,"perinuclear_halo":0.00,"multinucleation":0.00,"inflammation":0.00} Koilocytosis indicates atypical nuclei, Coarse chromatin indicates coarse chromatin granules, High N / C ratio indicates a high nucleocytoplasmic ratio, Hyperchromasia indicates deep nuclear staining, Nuclear irregularity indicates irregular nuclei, Perinuclear halo indicates a perinuclear halo, Multinucleation indicates multiple nuclei, and Inflammation indicates an inflammatory background. Example 2: Input description: The cells are generally regular in morphology, with no obvious abnormalities; some areas show a mild inflammatory background. Expected output: {"koilocytosis":0.00,"coarse_chromatin":0.00,"high_nc_ratio":0.00,"hyperchromasia":0.00,"nuclear_irregularity":0.00,"perinuclear_halo":0.00,"multinucleation":0.00,"inflammation":0.40} Example 3: Input description: Hollowed-out cells were found with perinuclear halos, suggesting HPV-related changes; The remaining features are atypical; Expected output: {"koilocytosis":0.70,"coarse_chromatin":0.10,"high_nc_ratio":0.10,"hyperchromasia":0.10,"nuclear_irregularity":0.10,"perinuclear_halo":0.75,"multinucleation":0.00,"inflammation":0.10} Example 4: Input description: Suspicious nuclear irregularity and suspicious nuclear deep staining, but the evidence is insufficient. It is recommended to combine more fields of view for evaluation. Expected output: {"koilocytosis":0.00,"coarse_chromatin":0.10,"high_nc_ratio":0.10,"hyperchromasia":0.40,"nuclear_irregularity":0.40,"perinuclear_halo":0.00,"multinucleation":0.00,"inflammation":0.00} Example 5: Input description: Multinucleated cells and a small amount of coarse granular chromatin are visible against an inflammatory background; Expected output: {"koilocytosis":0.00,"coarse_chromatin":0.50,"high_nc_ratio":0.00,"hyperchromasia":0.10,"nuclear_irregularity":0.10,"perinuclear_halo":0.00,"multinucleation":0.80,"inflammation":0.70} B. Construct instruction templates; The specific process is as follows: instruction: You will receive a cervical cytology description. Please return only the scores for the 8 attributes specified in the JSON schema, in the range [0,1], rounded to two decimal places; enter 0 for unknown or unmentioned attributes; do not add any explanatory text. User content (variables) (Description of this patch): describe: ; Please output JSON, where the keys must be in the following order: {"koilocytosis":x1,"coarse_chromatin":x2,"high_nc_ratio":x3,"hyperchromasia":x4,"nuclear_irregularity":x5,"perinuclear_halo":x6,"multinucleation":x7,"inflammation":x8} The output should only contain this one line of JSON; C. Establish scoring rules; the specific process is as follows: Strongly positive keywords: significant | obvious | highly | typical | large amount | severe; The score corresponding to strongly positive keywords is 0.9; Medium-level keywords: moderate | high | multiple | numerous; The corresponding score for medium-level keywords is 0.7; Mild keywords: mild | few | localized | occasional; Mild keywords correspond to a score of 0.5; Only mentioning words with no intensity; mentioning only words with no intensity corresponds to a score of 0.6; The score corresponding to the uncertain trigger word is based on the above score. ; The score corresponding to the negation trigger word is 0.0; High N / C Ratio Score Numerical hints: when At that time, the high nucleo-cytoplasmic ratio corresponds to a score of 0.95; when At that time, the high nucleo-cytoplasmic ratio corresponds to a fraction of 0.85; when At that time, the high nucleo-cytoplasmic ratio corresponds to a fraction of 0.75; D. Input A, B, and C into the large language model; the large language model outputs an attribute vector. ; Output attribute vectors for large language models Perform JSON parsing; if JSON parsing is successful, execute E; if JSON parsing fails, re-execute A, B, C, and D until JSON parsing is successful, then execute E. E. Attribute Vectors Each value is cropped to [0,1] and rounded to two decimal places to obtain the attribute vector formed by few-shot QA. .

8. The multi-turn dialogue-based intelligent diagnostic interactive system for cervical cell panoramic imaging according to claim 7, characterized in that: The context-aware multi-turn question-answering module is used to base on fused context representation. And the optimal instruction Prompt yields the final answer; The specific process is as follows: S41. Obtaining fusion context representation based on attention mechanism The specific process is as follows: S411. After each round of dialogue, the question and answer pair... Encoded by a text encoder into a fixed-dimensional history context vector; represented as: in, Indicating the first in history Question and answer pairs in turn-based dialogue Encoded context vector; Indicates a text encoder; Indicates the first in history The question in the round-robin dialogue; Indicates the first in history The response to the round-robin dialogue; This indicates vector concatenation; S412, When the user requests the current round New problems At that time, the new problems in the current round will be addressed. Input the text encoder to get a new question. ; New problems The set of vectors and historical vectors Input is a cross-attention mechanism, output is a historical context representation. ; indicates as: Among them, the historical vector set ; This indicates the question-and-answer session of the first round of historical dialogue. Encoded context vector; This indicates the question and answer session for the second round of historical dialogue. Encoded context vector; Indicating the first in history Question and answer pairs in turn-based dialogue Encoded context vector; S42. Obtaining the optimal instruction Prompt based on the new problem in S41; the specific process is as follows: The current round proposed by the user New problems Input a lightweight intent classifier, which will then process the user's new question. Mapped to predefined intent categories; represented as: in, Indicates will The input is the vector output by the text encoder; Indicates weight; Indicates bias; express Activation function layer The activation function layer outputs probability values; This indicates the category corresponding to the highest probability value; Indicates category; Select Category The corresponding instruction Prompt is the optimal instruction Prompt; S43, Historical Context Representation Based on S41 and the current round The image patch context generates the final answer; the specific process is as follows: S431. Construct the current round based on the retrieval enhancement generation mechanism. Image patch context; the specific process is as follows: Search enhancement generation mechanism: (1) When a user's question involves uncertain or rare cases, based on the currently viewed image patch. The context retrieves the most similar cases from a validated, annotated historical case database. One reference case; The specific process is as follows: Calculate the currently interested image patch in the Context Similarity to image patches in historical case database ; Calculate the similarity between the optimal instruction Prompt obtained in S42 of the Context and the instruction Prompts in the historical case database. ; Calculate the similarity between each augmentation information in the Context and augmentation information in the historical case database. ; Similarity Similarity Similarity The weighted average is then used to obtain the final similarity score. Select cases with high final similarity from the historical case database. One reference case; (2) Inject the prompts from the retrieved reference cases as additional context into the prompts in the Context; (3) Based on the Prompt instruction injected after (2) and the currently focused image block And optional enhanced information to construct the current round Image patch context; The optional enhancement information includes: category, source, triggering condition, injection method, and example; Category 1: Search the reference case database; Category 1 corresponds to the following source: annotated cervical pathology case database; The trigger condition for Category 1 is: the user's problem involves rare or uncertain features; The injection method corresponding to Category 1 is: concatenating the retrieved case report text into the Prompt; Reference cases retrieved for Category 1: Case A: Squamous epithelial cells with regular morphology and uniform staining: Diagnosis: NILM; Case B: A small number of atypical cells were observed against an inflammatory background, but not enough to diagnose LSIL. Diagnosis: ASC-US; Case C: Atypical nuclei and vacuolated cells were observed, consistent with the characteristics of a mild lesion. Diagnosis: LSIL; Case D: The nucleoplasm-cytoplasm ratio was significantly elevated, but some features were insufficient to completely confirm HSIL. Diagnosis: ASC-H; Case E: Nucleocytoplasmic ratio >0.8, deeply stained nuclei, coarse chromatin; Diagnosis: HSIL; Category 2: Medical Knowledge Base Entries; Category 2 corresponds to the following sources: WHO classification, Bethesda system, pathology textbooks or guidelines; the Bethesda system is the authoritative guideline for the diagnosis of cervical cancer. The triggering condition for Category 2 is: user issues involving terminology explanations or diagnostic criteria; The injection method corresponding to Category 2 is to inject the standard entries into the Prompt in the form of a short definition; The knowledge base reference for Category 2 is as follows: NILM: No intraepithelial lesions or malignant lesions were observed; ASC-US: Mild squamous epithelial cell abnormality, significance unclear; LSIL: Low-grade squamous intraepithelial lesion, often associated with HPV infection; ASC-H: Atypical squamous cells, HSIL cannot be ruled out; HSIL: High-grade squamous intraepithelial lesion, with an increased nuclear-to-cytoplasmic ratio, coarse or uneven chromatin, and an irregular nuclear membrane; Category 3: Statistics and Prior Information; Category 3 corresponds to the following source: historical case data; The triggering conditions for category 3 are: users asking probabilities and epidemiological questions; The injection method corresponding to category 3 is to supplement statistical data or incidence rates in the Prompt; The historical statistics for Category 3 are as follows: NILM: accounts for 85–90% of cervical smears; ASC-US: accounts for 5–7% of cervical smears; LSIL: accounts for 2–4% of cervical smears; ASC-H: 0.5–1% of cervical smears; HSIL: accounting for 1–2% of cervical smears; Category 4: Cross-modal hints; Category 4 corresponds to the following sources: ROI detection or attention heatmap; The trigger condition for category 4 is: user request for location or visualization support; The injection method corresponding to category 4 is to inject the ROI coordinates and heatmap descriptions into the Prompt. Model localization hint for category 4: Three abnormal cells were detected within the image patch. Cell 1: Normal nucleocytoplasmic ratio, regular morphology, corresponding to NILM; Cell 2: Mildly irregular, inflammatory background, corresponding to ASC-US; Cell 3: Increased nucleocytoplasmic ratio, coarse chromatin, corresponding to HSIL; Category 5: Longitudinal comparison of patient histories; Category 5 corresponds to the source: time-series slices from the same patient; The trigger condition for category 5 is: user questions involving disease progression; The injection method corresponding to category 5 is: injecting diagnostic comparison information from different time points into the Prompt; Patient follow-up information corresponding to category 5: March 2022: Diagnosed with NILM, which indicates normal cell morphology; June 2023: Diagnosed with ASC-US, which indicates the presence of a small number of atypical cells; January 2024: Diagnosed with LSIL, which indicates mild atypical nuclei; June 2025: Diagnosed with HSIL, which indicates an elevated nucleocytoplasmic ratio and coarse chromatin; S432, Represent the historical context of S41 The current wheel constructed with S431 The image patch context input S321 yields a fine-tuned Visual Large Language Model (VLM), which then generates the final answer. ; 。 9. The multi-turn dialogue-based intelligent diagnostic interactive system for cervical cell panoramic imaging according to claim 8, characterized in that: The final diagnostic suggestion generation module is used to calculate the average anomaly score based on the attribute vector, perform knowledge graph reasoning based on the average anomaly score, perform probabilistic model prediction based on the attribute vector, and obtain the final diagnostic suggestion based on knowledge graph reasoning and probabilistic model prediction. The specific process is as follows: S51, Attribute vectors based on all valid image patches Calculate the average outlier score for the entire WSI; expressed as: in, Indicates the first The average anomaly score of the entire WSI across all dimensions; Indicates the first Image blocks The weights; Represents attribute vector The dimension; S52, Attribute vector based on each image patch Calculate the attribute score for each image patch, and generate an interactive diagnostic heatmap covering the entire WSI based on the attribute score of each image patch; in, Indicates the first Attribute scores for each image patch; Indicates the first Weights of each dimension; S53. Knowledge graph reasoning based on average anomaly scores; the specific process is as follows: A knowledge graph was constructed based on authoritative guidelines for cervical cancer diagnosis. ; The average outlier score of the entire WSI As a query, in the knowledge graph The system performs reasoning to generate a diagnostic path. S54. Probabilistic model prediction based on attribute vectors; the specific process is as follows: Based on attribute vectors Constructing feature vectors ; eigenvectors Input a pre-trained probability prediction model, and the pre-trained probability prediction model outputs the posterior probability of different lesion grades; S55. A final diagnostic recommendation is obtained based on knowledge graph reasoning and probability model prediction; the specific process is as follows: The final diagnostic recommendation integrates the results of knowledge graph reasoning and probabilistic model prediction; The knowledge graph is as follows: 1) Construct the nodes and edges of the knowledge graph; specifically: 11) The node types of the knowledge graph are: Specimen and quality: Adequate, Inadequate, Obscuring blood / mucus, Inflammation; Adequate indicates acceptable; Inadequate indicates unacceptable / insufficient quality; Obscuring blood / mucus indicates blood / mucus obstruction. Inflammation indicates inflammation; Cytological characteristics: High N / C ratio, Hyperchromasia, Coarse chromatin, Nuclearirregularity, Koilocytosis, Multinucleation, Perinuclear halo, Mitosis, Parakeratosis, Keratinization, Inflammation; High N / C ratio indicates a high nucleocytoplasmic ratio; Hyperchromasia indicates deep nuclear staining; Coarse chromatin indicates coarse chromatin granules; Nuclear irregularity indicates nuclear irregularity; Koilocytosis indicates koilocytes or atypical nuclei; Multinucleation indicates multinucleated cells; Perinuclear halo indicates a perinuclear halo; Mitosis indicates atypical mitosis; Parakeratosis indicates parakeratosis; Keratinization indicates keratinization; Inflammation indicates inflammation. Composite patterns: Low-grade pattern, High-grade pattern, Reactive change; Low-grade pattern indicates a low-level pattern; High-grade pattern indicates a high-level pattern; Reactive change indicates a reactive change. Bethesda's diagnostic conclusions: NILM, ASC-US, LSIL, ASC-H, HSIL, SCC; NILM indicates negative; ASC-US indicates atypical squamous cells of indeterminate significance; LSIL indicates low-grade squamous intraepithelial lesion; ASC-H indicates atypical squamous cells, but HSIL cannot be ruled out; HSIL indicates high-grade squamous intraepithelial lesion; SCC indicates squamous cell carcinoma. Management recommendations: HPV typing, colposcopy, short-interval follow-up, expert review; HPV typing refers to HPV typing test; Colposcopy refers to colposcopy; Short-interval follow-up refers to short-term follow-up; Expert review refers to expert review. 12) The edge types of the knowledge graph are: Support or promote: Feature → Diagnosis; Feature → Diagnosis indicates that a certain case feature supports or reinforces a certain diagnosis; Conflict or Rejection: Pattern Diagnosis; Pattern Diagnosis indicates that a pattern conflicts with or is mutually exclusive with a diagnosis. Compositional relationship: Feature → Composite Pattern; Feature → Composite Pattern indicates that low-level features are combined together to form a pattern; Prerequisite: Diagnosis → {Essential Feature / Pattern}; Diagnosis → {Required Feature / Pattern} indicates that a diagnosis must possess certain essential features or patterns; Modulation relationship: Quality / Inflammation → Adjusting evidence weight; Quality / Inflammation → Adjusting the weight of evidence indicates that some background nodes in the graph are not direct diagnostic evidence, but will affect the weight of evidence. Management suggestion: Diagnosis → Management; Diagnosis → Management indicates that the diagnostic results directly trigger downstream management recommendations; 2) Construct the average outlier score for the entire WSI. Mapping to nodes; specifically: Input: The average outlier score for the entire WSI ; Mapping: Mapping each node With the corresponding attribute components Alignment, and by threshold With temperature parameters Perform smooth activation: Use the Sigmoid activation function; For node mapping functions; 3) Combining multiple features into composite patterns using fuzzy logic: High-level mode : in, For a high nucleus-to-cylinder ratio, For deep staining of the nucleus, It consists of crude nuclear chromatin; Low-level mode : in, It is an heterogeneous nucleus. Vacuole / hollowed-out cells; reactive or inflammatory pattern : in, Inflammation; 4) Rule Template: Define necessary / sufficient condition rules for each Bethesda diagnostic conclusion: Conditions for NILM: No significant abnormal features, both HGP and LGP are low; in, This indicates the score for the NILM diagnostic category; The criteria for ASC-US are: mild atypicality exists, but not enough to confirm LSIL; in, This indicates the score for the ASC-US diagnostic category; The criteria for LSIL are: low-grade HPV-related lesions, with Koilocytosis and Perinuclear halo as the core. in, This indicates the score for the LSIL diagnostic category; ASC-H criteria: Some features of high-grade lesions are present, but not enough to fully confirm HSIL; in, This indicates the score for the ASC-H diagnostic category; Criteria for HSIL: prominent high-grade lesion features; in, This indicates the score for the HSIL diagnostic category; "Reactive change" indicates a change in responsiveness. Squamous cell carcinoma (SCC): in, This indicates the score for the SCC diagnostic category; 5) Knowledge graphs provide edge weight patterns → Diagnosis; Pattern → Diagnosis indicates that a composite feature pattern conflicts with or is mutually exclusive with a certain diagnosis; The initial values ​​for the edge weights given by the knowledge graph are based on expert experience. 6) Quality control and inflammation modulation: If WSI quality is insufficient: the results are unreliable; If the inflammation score on the WSI is greater than 0.6, the confidence level for HSIL / ASC-H is down-adjusted, and the confidence level for ASC-US is up-adjusted. 7)、 71) The score based on the rule template in 4) is fused with the lower limit of the interval for each category in the historical statistical reference corresponding to category 3 in S431 to obtain the final diagnostic probability. ; The historical statistical reference for Category 3 includes the following intervals for each category: NILM: accounts for 85–90% of cervical smears; ASC-US: accounts for 5–7% of cervical smears; LSIL: accounts for 2–4% of cervical smears; ASC-H: 0.5–1% of cervical smears; HSIL: accounting for 1–2% of cervical smears; The lower limit of the interval for NILM is 85%; for ASC-US it is 5%; for LSIL it is 2%; for ASC-H it is 0.5%; and for HSIL it is 1%. Represented as: in, This represents the lower bound of the interval fusion category prior for each category in the historical statistical reference corresponding to category 3 in S431; This represents the score of rule template 4); Represents a monotonic function; Indicates the activation function; Indicates proportionality; This represents the average anomaly score for the entire WSI. Belongs to each category The probability of; 72) The average outlier score of the entire WSI. The input is a supervised learning model, and the output of the supervised learning model is the final diagnosis probability. ; 73) Probability of final diagnosis and final diagnosis probability Weighted fusion is performed to obtain the final diagnostic probability. ; indicates as: in, Indicates the probability of final diagnosis. The weights; Indicates the probability of final diagnosis. The weight.

10. The multi-turn dialogue-based intelligent diagnostic interactive system for cervical cell panoramic imaging according to claim 9, characterized in that: The multi-turn dialogue-based intelligent diagnostic interactive system for cervical cell panorama also includes an interactive and traceable diagnostic report output module. An interactive and traceable diagnostic report output module is used for multimodal information fusion and bidirectional link interface generation, as well as report generation; the specific process is as follows: S61. Multimodal information fusion and bidirectional link interface generation; the specific process is as follows: The user interface incorporates the following elements: (1) Global Heatmap View: The lowest resolution thumbnail in the full-slice image WSI The heatmap generated by S52 is superimposed on it; (2) Two-way link: Clicking on any highlighted area on the heatmap will jump to the corresponding image block view; at the same time, the text of the corresponding image block in S32 and the corresponding image in S32 will be displayed simultaneously on the right side of the user interface. piece The attribute vector corresponding to the text and the historical question and answer records of the corresponding image patch in S4; Conversely, clicking on the text of the corresponding image block in S32, the attribute vector corresponding to the text of the corresponding image block in S32, or the historical Q&A record of the corresponding image block in S4 will locate the corresponding image block. (3) Question and answer history stream: Retain the complete dialogue record of S1-S61; S62. Generate a report that combines structured and narrative elements; the specific process is as follows: A report that combines structured and narrative elements includes: Sample information: WSI ID: XXX, staining method: Pap smear; Overall slide quality assessment: Good, sufficient cell quantity, and uniform staining; High-risk area statistics: A total of N highly suspicious areas were found; Key areas represent images: clickable image patches Patch1, Patch2, and Patch3; Key pathological findings: Atypical cells with increased nuclear-cytoplasmic ratio, irregular nuclear membrane, and granular chromatin were detected in three image blocks, consistent with high-grade lesion characteristics; Diagnostic recommendation: High-grade squamous intraepithelial lesion [Confidence level: XX%]; Recommendations for auxiliary diagnosis: It is recommended that the diagnosis be reviewed and confirmed by a senior pathologist. If clinical conditions permit, HPV genotyping may be considered as supporting evidence.

Images