A deep learning-based cytological image rapid and accurate identification method and system

Abstract

The application discloses a kind of based on deep learning cytological image fast accurate identification method and system. Including: obtaining FNAC cytological image, and the image is preprocessed;Multi-scale representation fusion network model is constructed, and the FNAC cytological image after pre-processing is input into multi-scale representation fusion network model and is fused to obtain fusion feature;Fusion feature is input into edge dynamic enhancement algorithm and is handled to obtain enhanced feature;Enhanced feature is normalized and accessed long-range dependence modeling unit to integrate global feature, then through channel attention to effective channel empowerment purifies key features, to obtain final output feature;Final output feature is pooled and classified head output FNAC image's predicted class and confidence, and simultaneously generate saliency heat map to indicate key abnormal cell area.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and clinical diagnostic technology, and more specifically to a method and system for rapid and accurate identification of cellular images based on deep learning. Background Technology

[0002] Currently, existing deep learning models still have significant limitations in FNAC image recognition: traditional convolutional neural networks (CNNs) are limited by insufficient receptive fields, have low tolerance to low cell counts and complex background impurities, and are prone to problems such as insufficient feature expression and unstable class boundaries; at the same time, FNAC images generally have discrete cell distribution, background interference such as blood / inflammatory cells / mucus / debris, and significant morphological overlap between some lesions, which further exacerbates the recognition difficulty.

[0003] The core drawbacks of existing technologies are as follows: Insufficient specimen quality and sampling reliability: FNAC is highly sensitive to the operator's sampling technique and slide preparation quality, especially in specimens with "trace amounts of cells and low cell counts", which can easily lead to misdiagnosis, missed diagnosis or "insufficient specimen / inability to diagnose", resulting in repeated punctures or further biopsies, increasing the risk of secondary injury to patients and delays in diagnosis and treatment; Interpretation is highly subjective and lacks consistency: cytological morphology is complex and the manifestations of different lesions overlap, and the diagnostic results are highly dependent on the experience of the interpreter; different institutions have differences in classification standards, reporting norms and handling of "gray zone" cases, which limits the consistency and reproducibility of diagnosis and makes it difficult to meet the needs of accurate stratification and standardized management. Insufficient information utilization and lack of decision-making loop: Traditional processes rely heavily on single cytological images or fragmented information, lacking objective assessment and standardized output of sampling quality and diagnosticability. This makes it difficult to quickly generate risk stratification and treatment recommendations that directly correspond to clinical decisions, affecting triage efficiency and subsequent management.

[0004] Therefore, there is an urgent need for an intelligent diagnostic method that relies solely on a single FNAC image modality and balances accuracy and robustness to meet the needs of complex clinical scenarios. Summary of the Invention

[0005] In view of this, the present invention provides a method and system for rapid and accurate identification of cytological images based on deep learning. Using FNAC cytological images as the sole input, the method achieves accurate identification in complex scenes through a process of multi-scale feature extraction → fine-grained optimization of shallow features → multi-scale feature fusion → dynamic enhancement of fused feature edges → classification output.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for fast and accurate identification of cytological images based on deep learning, comprising: Step 1: Acquire FNAC cytology images and preprocess the images; Step 2: Construct a multi-scale representation fusion network model. Input the preprocessed FNAC cytology image into the multi-scale representation fusion network model for fusion processing to obtain fused features. Step 3: Process the fused features into the edge dynamic enhancement algorithm to obtain enhanced features; Step 4: Normalize the enhanced features and integrate them into the long-range dependency modeling unit to integrate global features. Then, use channel attention to weight and refine key features of effective channels to obtain the final output features. Step 5: The final output features are pooled and the classification head outputs the predicted class and confidence of the FNAC image, and a saliency heatmap is generated simultaneously to indicate key abnormal cell regions.

[0007] Preferably, step 2 includes: The preprocessed FNAC cytology image is input into a lightweight multi-scale feature extraction backbone network. The five-level features are obtained, and the formula is as follows: ; shallow features The fine-grained optimization features are obtained by inputting the Multidimensional Cosine Fine-Grained Optimization (MDFO) algorithm. ,right Smoothing and denoising and enhancing spatial representation are obtained Multi-scale representation fusion is achieved by using a step-by-step cascading approach with skip connections.

[0008] Preferably, the cascading method of the skip connection includes: Will and Tensor obtained by splicing tensor Fusion features are obtained through normalization and convolutional activation. and continue to work with each other in the same way. By merging step by step, fusion characteristics are obtained. .

[0009] Preferably, the fusion feature The input edge dynamic enhancement algorithm MGEA is given. First, Gaussian stabilization is performed, using the following formula: ; in, For Gaussian smoothing kernel, The standard deviation / smoothing scale of the Gaussian kernel; and the feature Convolution is used for noise reduction and feature stabilization; Then, a multi-directional edge operator is used to extract the high-frequency edge response, as shown in the following formula: ; in, For the first An edge operator in each direction is used to extract high-frequency edge responses; Then, the original features and high-frequency features are dynamically fused using learnable coefficients to obtain enhanced features: .

[0010] Preferably, the formula for obtaining the final output feature is as follows: ; in, The spatial dimensions of the feature map. For height, Width; The channel weight / attention coefficient vector is given by The result is obtained after two layers of linear transformation and activation, and is used to weight the feature channels; This is the Sigmoid activation function, used to generate gating / attention weights.

[0011] Preferably, a deep learning-based system for rapid and accurate identification of cellular images includes: Image acquisition module: used to acquire FNAC cytology images and preprocess the images; Feature fusion module: Used to construct a multi-scale representation fusion network model. The preprocessed FNAC cytology image is input into the multi-scale representation fusion network model for fusion processing to obtain fused features; Feature enhancement module: Used to process the fused features into the edge dynamic enhancement algorithm to obtain enhanced features; Feature extraction module: Normalizes the enhanced features and integrates them with the long-range dependency modeling unit to integrate global features. Then, it uses channel attention to weight and refine key features of effective channels to obtain the final output features. Output module: Used to pool and classify the final output features to output the predicted class and confidence of the FNAC image, and simultaneously generate a saliency heatmap to indicate key abnormal cell regions.

[0012] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method and system for rapid and accurate identification of cellular images based on deep learning, with the following beneficial effects: 1. Improved accuracy: Under the premise of using only FNAC image single modality, the accuracy of benign and malignant identification is ≥92%, and the accuracy of low cell quantity specimen identification is more than 15% higher than that of traditional CNN model, effectively reducing the false negative rate (≤3%) and false positive rate (≤5%). 2. Enhanced robustness: It can effectively suppress background interference such as blood, inflammatory cells, mucus, and debris, and improve the Kappa value of the consistency in identifying cases in the morphologically overlapping "gray area" from 0.65 to over 0.82, adapting to the differences in samples from different centers and with different staining methods; 3. Efficiency optimization: The model inference speed is ≤3 seconds / image, supporting rapid pre-screening of batch samples, which can shorten the image reading time by more than 60% and reduce the workload of physicians; 4. Strong clinical adaptability: The standardized output results are compatible with the existing risk stratification framework, and the heat map visualization assists in the review, making it easy to promote and apply in medical institutions at all levels, especially suitable for primary hospitals and scenarios where there is a lack of senior cytologists. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0014] Figure 1 The method flowchart provided by the present invention.

[0015] Figure 2 This is a structural schematic diagram provided for the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] The purpose of this invention is: Based solely on a single modal input of FNAC cytology images, this system enables automated and precise identification of common FNAC-indicated lesions such as thyroid nodules, enlarged lymph nodes, breast / salivary gland masses, and other punctureable lesions, improving the accuracy and stability of lesion nature (benign / malignant / inflammatory infection tendency) and specific type identification. For complex situations such as low cell count, background interference, and overlapping morphologies, the model automatically detects and focuses on analyzing key abnormal cells or suspected areas, reducing the probability of missed detection and misjudgment, and improving the clinical usability of trace cell specimens. It outputs standardized, reproducible results (including category, confidence level, risk indication, etc.), reducing reliance on individual experience, improving diagnostic consistency among different physicians and centers, and providing a clear basis for repeat punctures, further biopsies, or treatment decisions. like Figures 1-2 As shown in the figure, this invention discloses a method for fast and accurate identification of cytological images based on deep learning, including: Step 1: Acquire FNAC cytology images and preprocess the images; Step 2: Construct a multi-scale representation fusion network model. Input the preprocessed FNAC cytology image into the multi-scale representation fusion network model for fusion processing to obtain fused features. Step 3: Process the fused features into the edge dynamic enhancement algorithm to obtain enhanced features; Step 4: Normalize the enhanced features and integrate them into the long-range dependency modeling unit to integrate global features. Then, use channel attention to weight and refine key features of effective channels to obtain the final output features. Step 5: The final output features are pooled and the classification head outputs the predicted class and confidence of the FNAC image, and a saliency heatmap is generated simultaneously to indicate key abnormal cell regions.

[0018] Specifically, step 2 includes: The preprocessed FNAC cytology image is input into a lightweight multi-scale feature extraction backbone network. The five-level features are obtained, and the formula is as follows: ; shallow features The fine-grained optimization features are obtained by inputting the Multidimensional Cosine Fine-Grained Optimization (MDFO) algorithm. ,right Smoothing and denoising and enhancing spatial representation are obtained Multi-scale representation fusion is achieved by using a step-by-step cascading approach with skip connections.

[0019] In a specific embodiment of the present invention, to address the issue that shallow features are more sensitive to noise, the shallow features are first... Input MDFO to obtain fine-grained optimized features MDFO enhances discriminative cues through multidimensional feature rearrangement and interactive fusion, which can be exemplified as follows:

[0020] in, The fine-grained discriminative feature / weight map after multidimensional feature interaction fusion (obtained by Softmax normalization and multiplication of features rearranged in different dimensions) is used to highlight more discriminative cellular patterns and suppress irrelevant background features.

[0021] : No. There are 10 feature vectors to be determined.

[0022] : at the threshold The cost / number of false positives (misclassifying non-lesion / background as lesion, or misclassifying low risk as high risk).

[0023] (rho): Cost-weighting coefficient / trade coefficient, used to balance the importance of false positives and false negatives; The larger the value, the more optimization tends to reduce it. , The larger the value, the more likely it is to decrease. .

[0024] Then, using the reference vector Using this as a reference, calculate the feature vector. Cosine similarity:

[0025] Reference vector It is a reference feature vector obtained from the statistical analysis of background or non-lesion area features, used to compare with... Cosine similarity is used to measure the degree of similarity between the target and the background, thereby assisting in threshold discrimination and misclassification cost control.

[0026] And select or learn the threshold by minimizing the cost. To mitigate the cost of misclassification, it can be exemplified as follows: ; Furthermore, suppressing misclassification costs involves transforming the selection of thresholds from fixed empirical values ​​into controllable, learnable, and clinically optimal decisions, thereby more reliably controlling misclassifications in scenarios like FNAC, which involve "low cell count + complex background + overlapping morphology".

[0027] Specifically, the cascading method of the skip connection includes: Will and Tensor obtained by splicing tensor Fusion features are obtained through normalization and convolutional activation. and continue to work with each other in the same way. By merging step by step, fusion characteristics are obtained. .

[0028] In a specific embodiment of the present invention, after obtaining Then, it is smoothed and denoised, and its spatial representation is enhanced by deconvolution or upsampling to obtain... Subsequently, a cascading approach with skip connections is used to achieve multi-scale representation fusion: and Tensor obtained by splicing The fused features are obtained through normalization and convolutional activation. and continue in the same way with By fusing step by step, the fusion characteristics are obtained: .

[0029] Specifically, the fusion features The input edge dynamic enhancement algorithm MGEA is given. First, Gaussian stabilization is performed, using the following formula: ; in, For Gaussian smoothing kernel, The standard deviation / smoothing scale of the Gaussian kernel; and the feature Convolution is used for noise reduction and feature stabilization; Then, a multi-directional edge operator is used to extract the high-frequency edge response, as shown in the following formula: ; in, For the first An edge operator in each direction is used to extract high-frequency edge responses; Then, the original features and high-frequency features are dynamically fused using learnable coefficients to obtain enhanced features: .

[0030] Specifically, the formula for obtaining the final output feature is as follows: ; in, The spatial dimensions of the feature map. For height, Width; The channel weight / attention coefficient vector is given by The result is obtained after two layers of linear transformation and activation, and is used to weight the feature channels; This is the Sigmoid activation function, used to generate gating / attention weights.

[0031] In a specific embodiment of the present invention, in order to balance local edge enhancement and global contextual consistency, the following can be done: The data is normalized and connected to a long-range dependency modeling unit. Then, channel attention is used to weight the effective channels to obtain the final output features.

[0032] Long-range dependency modeling units obtain edge dynamic enhancements By inputting long-range dependency modeling units, scattered abnormal cell cues can be aggregated across regions in FNAC cytology images, and the interference of background pseudo-edges such as blood / mucus / debris can be suppressed by global context, thereby improving the recognition stability and accuracy under low cell quantity and complex background conditions.

[0033] Will The pooling and classification head outputs the predicted category and confidence level of the FNAC image (e.g., benign, malignant, inflammatory infection tendency, or preset multi-classification); and simultaneously generates a saliency heatmap to indicate key abnormal cell regions, thereby achieving stable enhancement and accurate identification in complex backgrounds and low cell count scenarios by relying solely on a single modality of FNAC image.

[0034] Specifically, a deep learning-based system for rapid and accurate identification of cellular images includes: Image acquisition module: used to acquire FNAC cytology images and preprocess the images; Feature fusion module: Used to construct a multi-scale representation fusion network model. The preprocessed FNAC cytology image is input into the multi-scale representation fusion network model for fusion processing to obtain fused features; Feature enhancement module: Used to process the fused features into the edge dynamic enhancement algorithm to obtain enhanced features; Feature extraction module: Normalizes the enhanced features and integrates them with the long-range dependency modeling unit to integrate global features. Then, it uses channel attention to weight and refine key features of effective channels to obtain the final output features. Output module: Used to pool and classify the final output features to output the predicted class and confidence of the FNAC image, and simultaneously generate a saliency heatmap to indicate key abnormal cell regions.

[0035] Furthermore, the key points (distinctions and advantages) of this invention are as follows: Single-modal adaptability: It only relies on FNAC cytology image input, without the need for additional clinical data or multimodal image fusion, adapting to clinical sample collection scenarios and lowering the application threshold; Dual-stage feature enhancement strategy: Innovatively introduces a dual-stage mechanism of "shallow feature MDFO optimization + fusion feature MGEA enhancement", which not only solves the problem of noise accumulation in shallow features, but also enhances key structural information, significantly improving the recognition robustness in low cell quantity and complex background. Multi-scale fusion optimization: Skip connections are used for step-by-step cascading fusion to fully preserve features from details to semantics across all levels, avoiding the class boundary instability problem caused by insufficient receptive field in traditional CNNs; Standardized output capability: The output results are aligned with the clinical risk stratification framework, while providing a heatmap visualization interpretation, ensuring both diagnostic consistency and traceability.

[0036] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0037] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for rapid and accurate identification of cytological images based on deep learning, characterized in that, include: Step 1: Acquire FNAC cytology images and preprocess the images; Step 2: Construct a multi-scale representation fusion network model. Input the preprocessed FNAC cytology image into the multi-scale representation fusion network model for fusion processing to obtain fused features. The preprocessed FNAC cytology image is input into a lightweight multi-scale feature extraction backbone network. The five-level features are obtained, and the formula is as follows: ； shallow features The fine-grained optimization features are obtained by inputting the Multidimensional Cosine Fine-Grained Optimization (MDFO) algorithm. ,right Smoothing and denoising and enhancing spatial representation are obtained Multi-scale representation fusion is achieved by using a step-by-step cascading approach with skip connections. MDFO enhances discriminative cues through multidimensional feature rearrangement and interactive fusion, as shown below: in, This is a fine-grained discriminative feature / weight map obtained after multi-dimensional feature interaction fusion. It is derived from features rearranged in different dimensions, normalized by Softmax, and then multiplied. Then, using the reference vector Using this as a reference, calculate the eigenvector. Cosine similarity: ， For the first A feature vector to be determined Select or learn the threshold by minimizing the cost. The cost of suppressing misclassification is expressed as: ； For the threshold is The cost / number of false positives at that time Cost weighting coefficient / trade-off coefficient; Step 3: Process the fused features into the edge dynamic enhancement algorithm to obtain enhanced features; Step 4: Normalize the enhanced features and integrate them into the long-range dependency modeling unit to integrate global features. Then, use channel attention to weight and refine key features of effective channels to obtain the final output features. The fusion features The input edge dynamic enhancement algorithm MGEA is given. First, Gaussian stabilization is performed, using the following formula: ； in, For Gaussian smoothing kernel, The standard deviation / smoothing scale of the Gaussian kernel; and the feature Convolution is used for noise reduction and feature stabilization; Then, a multi-directional edge operator is used to extract the high-frequency edge response, as shown in the following formula: ； in, For the first An edge operator in each direction is used to extract high-frequency edge responses; Then, the original features and high-frequency features are dynamically fused using learnable coefficients to obtain enhanced features: ； Step 5: The final output features are pooled and the classification head outputs the predicted class and confidence of the FNAC image, and a saliency heatmap is generated simultaneously to indicate key abnormal cell regions.

2. The method for rapid and accurate identification of cytological images based on deep learning according to claim 1, characterized in that, The cascading method of the skip connection includes: Will and Tensor obtained by splicing tensor Fusion features are obtained through normalization and convolutional activation. and continue to work with each other in the same way. By merging step by step, fusion characteristics are obtained. .

3. The method for rapid and accurate identification of cytological images based on deep learning according to claim 1, characterized in that, The formula for obtaining the final output feature is as follows: ； in, The spatial dimensions of the feature map. For height, Width; The channel weight / attention coefficient vector is given by The result is obtained after two layers of linear transformation and activation, and is used to weight the feature channels; The Sigmoid activation function is used to generate gating / attention weights; To normalize the enhanced features and integrate them into the long-range dependency modeling unit to incorporate global features.

4. A rapid and accurate cytological image recognition system based on deep learning, used to implement the method described in any one of claims 1-3, characterized in that, include: Image acquisition module: used to acquire FNAC cytology images and preprocess the images; Feature fusion module: Used to construct a multi-scale representation fusion network model. The preprocessed FNAC cytology image is input into the multi-scale representation fusion network model for fusion processing to obtain fused features; Feature enhancement module: Used to process the fused features into the edge dynamic enhancement algorithm to obtain enhanced features; Feature extraction module: Normalizes the enhanced features and integrates them with the long-range dependency modeling unit to integrate global features. Then, it uses channel attention to weight and refine key features of effective channels to obtain the final output features. Output module: Used to pool and classify the final output features to output the predicted class and confidence of the FNAC image, and simultaneously generate a saliency heatmap to indicate key abnormal cell regions.

Images