Blood smear image segmentation and intelligent interactive diagnosis system based on deep learning

Abstract

The invention discloses a blood smear image segmentation and intelligent interactive diagnosis system based on deep learning, uses a mobile terminal, a cloud terminal and a PC terminal, belongs to the technical field of intelligent diagnosis, and comprises a mobile terminal acquisition and quality inspection module, a cloud reasoning platform, a PC terminal GUI and a mobile terminal result echo; the cloud reasoning platform comprises an o-plasmodium detection subsystem, an o-leukocyte classification / segmentation subsystem and an o-fusion and statistics subsystem, and the o-plasmodium detection subsystem is used for outputting candidate regions and confidence coefficients on the basis of improved AlexNet; and the leukocyte classification / segmentation subsystem is used for completing WBC five classification and mask generation based on the improved GoogLeNet. The diagnosis system converges a detection result, a mask, a candidate graph and a purification graph into a set of GUI linked on the same screen, parameter snapshot playback and candidate priority rechecking are introduced, the rechecking burden and parameter adjustment cost are remarkably reduced, traceability and consistency are improved, and by combining with an end-cloud-end closed loop, the diagnosis efficiency is improved. Compared with cloud reasoning and simple echoing or a single-machine processing tool, the method has obvious advantages.

Description

Technical Field

[0001] This invention belongs to the field of intelligent diagnostic technology, and in particular relates to a blood smear image segmentation and intelligent interactive diagnostic system based on deep learning. Background Technology

[0002] Traditional manual microscopy is an important method for malaria diagnosis, but it is highly dependent on the experience of the examiner, inefficient, and prone to missed or false positives in cases of low parasite density, uneven staining, or focusing errors. While RDT rapid test strips are portable, they are susceptible to false negatives due to antigen variation and sample quality. With the development of deep learning, detection and classification models for blood smears have emerged in academia and industry. However, common problems in engineering implementation include: 1. Fragmented interface: The original image, candidate blocks, mask, and cleanup image are displayed separately, resulting in poor spatial consistency and high review costs; 2. Weaknesses in auditing: Adjustments to parameters such as thresholds, NMS, and morphology cannot be replayed or compared; 3. Insufficient edge-cloud-edge collaboration: Mobile data collection, cloud inference, and PC verification lack consistency and fault tolerance during network outages; 4. Incomplete chain of evidence in the report: Statistical data and evidence (keyframes, parameters) are difficult to solidify with a single click.

[0003] Existing technology: Option A: Standalone image processing tool It only performs enhancement / threshold segmentation / counting; it lacks deep learning detection, multi-view linkage, and reporting evidence chains.

[0004] Option B: Cloud-based inference + simple app feedback Photos are uploaded via mobile phone, and the cloud returns positive / negative results and a small number of positive bounding boxes. The PC-based verification capability is weak, lacks candidate queue sorting, and has no parameter snapshots.

[0005] Option C: Academic Prototype / Training Script It provides AlexNet / GoogLeNet training and offline inference, but lacks a clinically-oriented GUI and auditing mechanism.

[0006] Existing shortcomings: Current intelligent blood smear diagnostic systems generally suffer from fragmented multi-view information at the interface level, making it difficult to achieve simultaneous screen interaction and consistent spatial positioning; parameter adjustment lacks replayable "parameter-result" snapshots and a complete audit chain, making review and traceability difficult; candidate region review does not combine confidence level and contextual comprehensive scoring, and also lacks efficient batch operation capabilities; the end-to-cloud-to-end process does not adequately support offline scenarios, and lacks task queues and result consistency merging mechanisms; at the same time, report generation cannot solidify image evidence and parameter metadata with one click, resulting in an incomplete evidence chain and insufficient clinical acceptance.

[0007] It should be noted that the above content falls within the inventor's technical knowledge and does not necessarily constitute prior art. Summary of the Invention

[0008] To address the aforementioned issues, the present invention aims to provide a deep learning-based blood smear image segmentation and intelligent interactive diagnostic system. This system constructs a multi-view linked graphical user interface and interaction method, enabling consistent presentation and linked operation of the original image, candidate regions, white blood cell regions, parasite masks, and cleaned images within the same interface. It establishes a "parameter-result" snapshot playback mechanism covering key parameters such as thresholds, NMS, and morphology, along with a candidate priority review process based on comprehensive scoring. Within an end-cloud-end collaborative framework, it establishes a closed loop connecting "mobile quality inspection and collection—cloud inference—PC review and reporting," providing network outage tolerance, task queuing, and result consistency merging capabilities. It also supports one-click import of evidence and parameter metadata into structured reports, thereby improving review efficiency and traceability, reducing false positives and false negatives, and adapting to grassroots application scenarios.

[0009] To achieve the above objectives, this invention proposes a blood smear image segmentation and intelligent interactive diagnostic system based on deep learning. The system can be used on mobile devices, cloud devices, and PC devices. The diagnostic system includes a mobile device acquisition and quality inspection module, a cloud inference platform, a PC GUI, a mobile device result display module, and an end-to-cloud interface module.

[0010] Cloud-based inference platforms, including: o Malaria parasite detection subsystem: Based on improved AlexNet output candidate regions and confidence scores.

[0011] o White blood cell classification / segmentation subsystem: Based on the improved GoogLeNet, complete the five-class classification of WBC and mask generation.

[0012] o Fusion and Statistics: Used to perform NMS and IoU filtering on candidate regions, calculate comprehensive scores by combining leukocyte context, output positive counts and concentration estimates, and return candidate regions, masks, thumbnails and statistics.

[0013] The improved AlexNet / GoogLeNet is a network with lightweight modifications added to the classic CNN structure; it is used for parasite detection and white blood cell classification in the diagnostic system.

[0014] Furthermore, the mobile acquisition and quality inspection module is used to connect to a microscope accessory to acquire blood smear images, perform brightness / focus quality inspection and lightweight enhancement, and encapsulate the images and metadata into JSON for uploading to the cloud inference platform.

[0015] Furthermore, the PC-side GUI system includes a multi-view linkage presentation module, a parameter management and snapshot module, a candidate queue management module, a report generation module, a mask overlay and side-by-side comparison module, and a log and security module.

[0016] The multi-view linkage rendering module is responsible for multi-view linkage rendering, the parameter management and snapshot module is responsible for parameter adjustment and snapshots, the candidate queue management module is responsible for candidate region verification, the report generation module is responsible for report generation and export, the mask overlay and side-by-side comparison module is used for evidence verification and false detection identification, and the log and security module is used for access control, de-identification and hash verification.

[0017] Furthermore, the mobile terminal result display is used to overlay the core results.

[0018] Furthermore, the PC-side GUI view includes a parameter and task panel, a main display area, and a multi-view stack.

[0019] Furthermore, the parameters and task panel includes image selection, confidence threshold, NMS and IoU, target size and morphological kernel, running detection, saving results, and launching the AI ​​assistant.

[0020] Furthermore, the main display area includes a detection result view, a toolbar, a detection box, and detection information; The detection results view is used to display the overall detection results of Plasmodium. It can overlay the Plasmodium mask and the detection box. At the same time, the main display area displays a summary of the detection information and provides patient-readable instructions. The toolbar supports zooming, side-by-side comparison and export. The detection box is used to outline the area containing Plasmodium when detecting a target, and is a local magnified view of the detection results view.

[0021] Furthermore, the multi-view stack includes a leukocyte gallery, a parasite candidate queue, and a mask and purification image.

[0022] Furthermore, the mask and purification map includes a parasite mask 1 and a leukocyte-removing map.

[0023] Furthermore, the linkage rules of the PC-side GUI include: When hovering over / clicking the parasite candidate queue, the overall detection results view is simultaneously positioned and highlighted; The mask and cleaned image support transparency sliders and side-by-side comparison of the original and cleaned images; In the PC GUI, all views share coordinates and zoom and translate synchronously. PC GUI Parameters - Result Snapshot Replay: Each parameter change automatically generates a parameter snapshot, supports one-click rollback and pairwise comparison, and writes it to the case audit log; The review of candidate priorities for the PC GUI includes: Overall score: Score = α·Conf + β·Ctx Where Conf represents the model confidence score; Ctx represents the context score. Candidate priority review supports batch approval, rejection, and re-examination. Approved items are entered into the evidence panel and used for reporting. Concentration estimation and statistics for PC-based GUI, including: The main display area shows the estimated concentration of malaria parasites, positive counts, and confidence intervals. By default, it uses the number of positive candidates per unit field of view / number of reference cells or uses area normalization for calculation. The coefficients are configurable.

[0024] The AI ​​assistant in the PC GUI and disclaimer include: Summarize current evidence and precautions to generate a "DetectionInformationSummary", clearly stating that it is "for reference only and cannot replace diagnosis by medical personnel".

[0025] The blood smear image segmentation and intelligent interactive diagnostic system based on deep learning proposed in this invention can bring the following beneficial effects: 1. The diagnostic system of this invention integrates the detection results, mask, candidates and cleanup images into a set of GUIs that are linked on the same screen, and introduces parameter snapshot playback and candidate priority review, which significantly reduces the review burden and parameter tuning cost, improves traceability and consistency, and combines end-to-end closed loop (including network outage fault tolerance and structured reports), which has obvious advantages over "cloud inference + simple display" or "standalone processing tools" in terms of clinical usability, audit compliance and engineering implementation. 2. The diagnostic system of this invention adopts an improved AlexNet and GoogLeNet model. Through lightweight modifications such as the introduction of an attention mechanism, the model's ability to extract features from Plasmodium microsporum and complex-shaped white blood cells is enhanced, significantly improving the accuracy and recall of candidate region localization and mask segmentation. The design of parallel inference of a dual-system (Plasmodium detection + white blood cell segmentation / classification) ensures that the system can comprehensively capture key information in the image, laying a solid foundation for subsequent fusion and contextual analysis.

[0026] 3. The diagnostic system of this invention adopts a multi-stage filtering and fusion strategy: by filtering through NMS and IoU and combining it with white blood cell context information (such as distance and background uniformity) to calculate a comprehensive score, a large number of false positive detection boxes caused by impurities and staining residues are effectively eliminated, which greatly improves the reliability of diagnosis. The "clean image" assists in judgment: it provides an image view after removing white blood cells, which can effectively highlight the suspected pathogen area and provide medical personnel with a unique tool to identify false detections and verify evidence.

[0027] 4. The diagnostic system of this invention not only provides positive counts, but also estimates concentrations by normalizing per unit field of view or by using reference cell counts, and provides confidence intervals, making the output results more clinically valuable and assisting doctors in assessing the severity of the condition. Attached Figure Description

[0028] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the structure of a deep learning-based blood smear image segmentation and intelligent interactive diagnostic system according to the present invention.

[0029] Figure 2 This is a schematic diagram of the improved GoogLeNet model of this invention; Figure 3 This is a flowchart illustrating the process of identifying malaria parasites using the AlexNet model in this invention. Figure 4 This is a schematic diagram of the PC-side GUI of a deep learning-based blood smear image segmentation and intelligent interactive diagnostic system according to the present invention.

[0030] In the diagram: 1. Parameters and Tasks Panel; 11. Image Selection; 12. Confidence Threshold; 13. NMS and IoU; 14. Target Size and Morphological Kernel; 15. Run Detection; 16. Save Results; 17. Start AI Assistant; 2. Main Display Area; 21. Detection Results View; 22. Toolbar; 23. Detection Information; 24. Detection Box; 3. Multi-View Stack; 31. Leukocyte Gallery; 32. Parasite Candidate Queue; 33. Mask and Cleanup Image; 331. Parasite Mask; 332. Leukocyte-Removed Image.

[0031] Figure 3 In Chinese: *X*X* (e.g., 227X227X3) is the H×W×C notation for tensor / image size, representing height*, width*, and * channels (RGB). For example, 227X227X3 represents a height of 227, a width of 227, and 3 channels. Detailed Implementation

[0032] To more clearly illustrate the overall concept of the present invention, a detailed description will be provided below with reference to the accompanying drawings and examples.

[0033] Embodiments of this invention propose a deep learning-based blood smear image segmentation and intelligent interactive diagnostic system, applicable to mobile devices, cloud-based systems, and PCs, such as... Figure 1As shown, the diagnostic system includes a mobile data acquisition and quality inspection module, a cloud-based inference platform, a PC-based GUI (graphical user interface), a mobile-based result display module, and an end-to-cloud interface module.

[0034] The mobile acquisition and quality inspection module is used to connect to a microscope accessory to acquire images of blood smears (samples that are smeared on a glass slide and stained before being observed under a microscope), perform brightness / focus quality inspection and lightweight enhancement, and encapsulate the images and metadata into JSON (a lightweight data exchange format used by the system-cloud communication module to return candidate regions, masks and statistics) and upload them to the cloud inference platform.

[0035] Cloud-based inference platforms, including: o Plasmodium (malaria-causing parasite) detection subsystem: Based on an improved AlexNet (a convolutional neural network model), it outputs candidate regions and confidence scores; o White blood cell classification / segmentation subsystem: Based on the improved GoogLeNet (a convolutional neural network model), it completes WBC five-class classification and mask generation; In the field of digital image processing, masking is a powerful method. Its core idea is to extract or highlight specific parts of an image through local region operations. It is commonly used in tasks such as image preprocessing, feature extraction, and image segmentation.

[0036] The improved AlexNet / GoogLeNet is a network with lightweight modifications (such as attention and channel enhancement) added to the classic CNN (convolutional neural network) structure; it is used in this diagnostic system for parasite detection / white blood cell classification.

[0037] Improved GoogLeNet, such as Figure 2 As shown, it specifically includes: I. Inception module optimization: Improvements were made to the original GoogLeNet model based on its four branches (ψ1–ψ4): ψ1 is the same as the original, that is, a 1×1 convolution, used for channel compression; ψ2 is the same as the original, that is, 1×1 convolution → 3×3 convolution, which is used to obtain local spatial features; ψ3 improves the original 1x1 convolution → 5x5 convolution to 1×1 convolution → 5×1 convolution → 1×5 convolution, so that factorized convolution can obtain slender morphological features; ψ4 improves the original 3x3 max pooling followed by 1x1 convolution to 3×3 max pooling followed by 3×3 convolution, in order to introduce pooling information; The gradient flow is improved by using a 1×1 residual channel, and the weight of the 1×1 residual channel is denoted as... Among them, ψ1–ψ4 are the four branches inside the Inception3, 4 and 5 modules in the GoogLeNet model.

[0038] II. Combination of Multiscale and Attention Mechanisms After the input is processed by multi-scale Gaussian convolution, attention modules SE (channel attention), CBAM (hybrid attention), and ECA (efficient channel attention) are introduced to weight the multi-scale features, thereby improving the model's sensitivity to subtle cell morphology.

[0039] III. Auxiliary Classifiers and Dynamic Loss Weights Two auxiliary classifiers are set in the middle of the network, and the ratio of main / auxiliary classification loss is adjusted by adaptive dynamic weights to enhance the stability of early training and improve the convergence speed.

[0040] IV. Optimized Training Scheduling Improved cosine annealing learning rate scheduling: periodic restarts and cycle-by-cycle reduction of the maximum learning rate make later refinement smoother.

[0041] V. Biological Priors and Data Augmentation Improvements Biological morphological constraints based on area, perimeter, and compactness are added to the loss function; at the same time, a Mixup strategy based on structural similarity (SSIM) is introduced into the data augmentation, which only performs Mixup on samples with low similarity and performs color consistency correction.

[0042] Improved AlexNet, such as Figure 3 As shown, it specifically includes: I. Input Terminal Improvement Candidate ROI screening and standardization are introduced at the input end, including white blood cell removal, overstained region filtering, and adaptive histogram equalization, to ensure that the input fed into the network is cleaner and to reduce interference from irrelevant regions.

[0043] ROI screening involves generating suspected target regions, deduplicating (NMS), and sorting them to obtain a candidate set to be reviewed.

[0044] II. Convolutional Layer Structure Optimization The first two convolutional layers use large convolutional kernels (11×11 convolution in the input layer and 5×5 convolution in the C1 layer) combined with LRN (cross-channel normalization, suppressing abnormally large activations, used in traditional AlexNet) and 3×3 convolutional overlapping max pooling in the C2 layer to quickly capture coarse features and stably normalize them; the middle and later convolutional layers (C3, C4, C5) use stacked 3×3 small convolutional kernels to extract detailed features.

[0045] III. Enhancement of Fully Connected Layers Two large-scale fully connected layers (FC6 and FC7, both 4096-dimensional) are used, along with ReLU activation and 0.5 Dropout (a regularization technique to prevent overfitting of neural networks), combined with L2 weight decay to effectively prevent overfitting.

[0046] L2 weight decay is a commonly used regularization technique during training. It adds an extra "penalty based on weight size" when calculating the loss, proportional to the sum of the squares of the parameters in each layer. This results in an additional "pulling back" adjustment to zero for the weights with each parameter update, preventing excessive weights, reducing overfitting, and improving generalization. The decay strength is controlled by the coefficient λ, typically ranging from 1e-5 to 1e-3. With standard SGD (Stochastic Gradient Descent), this is equivalent to adding a term proportional to the weights to the existing gradient. With optimizers like AdamW, a "decoupled" approach is usually used, applying weight decay and adaptive gradients separately. It's important to emphasize that L2 weight decay only takes effect during the training phase and does not change the computational flow or results during the inference phase.

[0047] IV. Improvement of Training Strategies Transfer learning is employed with a tiered learning rate setting: a small learning rate at the bottom and a large learning rate at the top; periodic cosine annealing is also introduced and the peak learning rate is reduced periodically to accelerate convergence.

[0048] o Fusion and Statistics: Perform NMS (Non-maximum suppression, used to remove highly overlapping detection boxes) and IoU (Intersection over Union, an indicator of the degree of overlap between two detection boxes) filtering on candidate regions; calculate a comprehensive score by combining white blood cell context, and output positive count and concentration estimate; return candidate regions, masks, thumbnails and statistics.

[0049] The PC-based GUI system includes a multi-view interactive presentation module, a parameter management and snapshot module, a candidate queue management module, a report generation module, a mask overlay and side-by-side comparison module, and a log and security module.

[0050] The multi-view linkage rendering module is responsible for multi-view linkage rendering; the parameter management and snapshot module is responsible for parameter adjustment and snapshots; the candidate queue management module is responsible for candidate area verification; the report generation module is responsible for report generation and export; the mask overlay and side-by-side comparison module is used for evidence verification and false detection identification; and the log and security module is used for access control, de-identification, and hash verification.

[0051] The mobile results display uses an overlay to show the core results. The overlay is a transparent / semi-transparent independent rendering layer that is placed on top of the original image view. It is responsible for visualizing the detection results (such as detection boxes, segmentation masks, label text, confidence scores, scales, heatmaps, etc.) without modifying the original image itself, which allows for quick review. When switching on / off, adjusting the transparency, or highlighting a candidate area, only the overlay needs to be redrawn, which is both lightweight and intuitive.

[0052] The PC GUI view includes the parameter and task panel 1, the main display area 2, and the multi-view stack 3, such as Figure 4 As shown.

[0053] The parameters and tasks panel 1 includes image selection 11, confidence threshold 12, NMS (non-maximum suppression) and IoU (intersection over union) 13, target size and morphological kernel 14, run detection 15, save results 16, and launch AI assistant 17.

[0054] Functions of NMS: When the model generates multiple highly overlapping candidate regions near the same target, NMS is used to "deduplicate". Specifically, it sorts the candidate regions from high to low confidence, keeps the box with the highest confidence, and deletes other boxes whose overlap (measured by IoU) with it exceeds a set threshold. Then, it repeats this process for the remaining candidate regions. The purpose is to eliminate duplicate boxes, improve readability and statistical accuracy, and prevent a target from being counted repeatedly. The function of IoU: IoU is used to measure the degree of overlap between two candidate regions, with a value ranging from 0 to 1. The IoU slider is used to provide the NMS with a threshold for determining whether two regions are considered the same target. In this interface, it mainly controls the overlap determination of the NMS and, together with the confidence threshold, determines the number and stability of the candidate regions that are ultimately retained. The main display area 2 includes a detection result view 21, a toolbar 22, a detection frame 24, and detection information 23. The detection results view 21 is used to display the overall detection results of Plasmodium. The Plasmodium mask and detection box 24 can be superimposed. At the same time, the main display area 2 displays a summary of detection information and provides patient-readable instructions. The toolbar 22 supports zooming, side-by-side comparison and export. The detection box 24 is used to outline the area containing Plasmodium when detecting a target, and is a local magnified view of the detection results view 21.

[0055] The multi-view stack 3 includes a leukocyte gallery 31, a parasite candidate queue 32, and a mask and purification image 33, which includes a parasite mask 331 and a leukocyte-removed image 332.

[0056] PC GUI linkage rules: When hovering / clicking on candidate parasite 32, the overall detection results view is simultaneously positioned and highlighted; Mask and cleanup image 33 supports transparency sliders and side-by-side comparison of original and cleanup images.

[0057] In the PC GUI, all views share coordinates and zoom and pan synchronously.

[0058] PC GUI Parameters - Result Snapshot Replay: Each parameter change automatically generates a parameter snapshot (parameter vector, thumbnail preview, input image hash, and timestamp), supports one-click rollback and pairwise comparison, and writes it to the case audit log.

[0059] Candidate priority review via PC GUI (ranked by a comprehensive score based on confidence level × context score, supporting batch approval, rejection, and re-review): Overall score: Score = α·Conf + β·Ctx Where Conf is the model confidence score; Ctx is the context score (such as the score for distance from WBC (white blood cells), background uniformity, local contrast, etc.).

[0060] Candidate priority review supports batch approval / rejection / re-examination. Approved items are entered into the evidence panel and used for reporting.

[0061] Concentration estimation and statistics using PC-based GUI: The main display area 2 shows the estimated concentration of malaria parasites, positive counts, and confidence intervals. The default calculation method is to use the number of positive candidates per unit field of view / number of reference cells or to use area normalization. The coefficients are configurable.

[0062] AI Assistant in PC GUI and Disclaimer: Summarize current evidence and precautions to generate a "Detection Information Summary," clearly stating that it is "for reference only and cannot replace diagnosis by medical personnel."

[0063] The method of using this diagnostic system includes the following steps: S1: Mobile sample collection and cloud-based intelligent initial screening, specifically including: S11: Sample Collection and Image Upload Medical staff obtain digital images of stained blood smears using microscope accessories; The mobile acquisition and quality inspection module automatically performs focus sharpness and brightness uniformity checks on the images. If the quality inspection passes, it performs lightweight image enhancement (such as contrast stretching) to optimize subsequent analysis.

[0064] The mobile acquisition and quality inspection module encapsulates images (Base64 encoded JPEG format) and metadata (such as case ID, device ID, magnification, focus score, etc.) into a predefined JSON format and uploads them to the cloud inference platform through a secure end-to-cloud interface.

[0065] S12: Cloud-based Parallel Inference and Fusion: The Plasmodium detection subsystem uses an improved AlexNet model to infer from images, generating candidate bounding boxes for suspected Plasmodium and their corresponding confidence scores.

[0066] The white blood cell classification / segmentation subsystem uses an improved GoogLeNet model to infer images, completes the five classifications of white blood cells (such as neutrophils, lymphocytes, etc.), and generates accurate pixel-level segmentation masks.

[0067] S13: Fusion and Statistics Module: Perform NMS and IoU-based filtering on candidate bounding boxes to remove highly overlapping false positive boxes.

[0068] The comprehensive score for each candidate region is calculated by combining the detected white blood cell context information (such as the distance between the malaria parasite and the white blood cell, the background environment, etc.).

[0069] The number of positive candidate regions was counted, and the concentration of malaria parasites was estimated based on the reference white blood cell count or field area.

[0070] The cloud-based inference platform encapsulates the final results (including candidate bounding boxes, malaria parasite and leukocyte masks, thumbnails, and statistics) into JSON format and returns them to PC and mobile devices.

[0071] S2: Interactive PC-based refinement and review diagnostics, specifically including: S21: Result Initialization and Multi-View Rendering: The PC-side GUI receives and parses the data returned from the cloud.

[0072] The main display area 2 renders the original blood smear image and overlays all malaria parasite detection boxes and masks with confidence levels higher than the initial threshold by default.

[0073] Multi-view stack 3 loads multiple auxiliary views simultaneously.

[0074] White blood cell gallery 31: Displays all identified white blood cells and their classification results.

[0075] Parasite candidate queue 32: All candidate targets are displayed in list form and sorted by overall score.

[0076] Figure 33: Showing images of parasite masking and “cleaning” after the removal of white blood cells, used to highlight suspected pathogens.

[0077] S22: Parameter Adjustment and Intelligent Interaction Medical staff can adjust parameters such as confidence threshold 12, NMS and IoU 13, and morphological kernel size in the Parameters and Tasks panel 1 in real time to change the screening strictness of candidate targets and mask morphology. The diagnostic system will then redraw all views immediately.

[0078] Linkage rule: When the mouse hovers over or clicks on a target in the parasite candidate queue 32, the view in the main display area 2 will automatically pan and zoom to the target and highlight it, achieving precise positioning.

[0079] Side-by-side comparison: Medical staff can use the toolbar function to compare the original image, overlaid mask image, and cleaned image side-by-side to assist in identifying false positives and verifying evidence.

[0080] S23: Candidate Review and Evidence Locking: Medical staff can review the candidate list sorted by comprehensive score (Score=α·Conf+β·Ctx), focus on high-scoring targets, and verify them one by one with the help of the multi-view interactive presentation module.

[0081] For targets of doubt, the AI ​​assistant can be invoked to obtain the basis for the model's judgment.

[0082] Medical staff can select candidate targets in batches and perform "pass", "reject" or "mark for review" operations. Only candidates that are "passed" will be finally confirmed as positive and used to generate reports.

[0083] S24: Generating and exporting diagnostic reports: After confirming all positive targets, click "Save Results 16".

[0084] The report generation module will automatically create a structured report, which includes: Keyframe: Evidence image with detection boxes and masks.

[0085] Statistical tables: Plasmodium positive count, concentration estimate (with confidence interval), and white blood cell differential count.

[0086] Parameter metadata: A snapshot of the parameters of the algorithm used, ensuring that the results are reproducible.

[0087] Evidence Puzzle: A collection of thumbnails of all positive targets.

[0088] Disclaimer: The test information is clearly marked as "for reference only and should not replace diagnosis by medical professionals".

[0089] Reports can be exported as PDFs or in formats supported by the Hospital Information System (HIS) and archived. The logging and security modules record all operations to ensure data security and auditability.

[0090] S3: Mobile Results Display and Archiving: Diagnostic reports and core results (such as positive / negative conclusions and key evidence diagrams) will be transmitted back to the mobile device for result display via the end-to-cloud interface.

[0091] Medical staff can quickly view diagnostic conclusions on mobile devices in the form of overlays, conduct bedside verification or simple communication, and complete the diagnostic closed loop from collection to verification.

[0092] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0093] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A blood smear image segmentation and intelligent interactive diagnostic system based on deep learning, using mobile devices, cloud-based devices, and PCs, characterized in that... The diagnostic system includes a mobile terminal data acquisition and quality inspection module, a cloud-based inference platform, a PC-based GUI, a mobile terminal result display module, and an end-to-cloud interface module. Cloud-based inference platforms, including: o Malaria parasite detection subsystem: Based on improved AlexNet output candidate regions and confidence scores; o White blood cell classification / segmentation subsystem: Based on the improved GoogLeNet, complete the five-class classification of WBC and mask generation; o Fusion and Statistics: Used to perform NMS and IoU filtering on candidate regions, calculate comprehensive scores by combining leukocyte context, output positive counts and concentration estimates, and return candidate regions, masks, thumbnails and statistics; The improved AlexNet / GoogLeNet is a network with lightweight modifications added to the classic CNN structure; it is used for parasite detection and white blood cell classification in the diagnostic system.

2. The blood smear image segmentation and intelligent interactive diagnostic system based on deep learning according to claim 1, characterized in that, The mobile acquisition and quality inspection module is used to connect to a microscope accessory to acquire blood smear images, perform brightness / focus quality inspection and lightweight enhancement, and encapsulate the images and metadata into JSON for uploading to the cloud inference platform.

3. The blood smear image segmentation and intelligent interactive diagnostic system based on deep learning according to claim 2, characterized in that, The PC-based GUI system includes a multi-view linkage presentation module, a parameter management and snapshot module, a candidate queue management module, a report generation module, a mask overlay and side-by-side comparison module, and a log and security module. The multi-view linkage rendering module is responsible for multi-view linkage rendering, the parameter management and snapshot module is responsible for parameter adjustment and snapshots, the candidate queue management module is responsible for candidate region verification, the report generation module is responsible for report generation and export, the mask overlay and side-by-side comparison module is used for evidence verification and false detection identification, and the log and security module is used for access control, de-identification and hash verification.

4. The blood smear image segmentation and intelligent interactive diagnostic system based on deep learning according to claim 3, characterized in that, The mobile terminal result display is used to overlay and display the core results.

5. The blood smear image segmentation and intelligent interactive diagnostic system based on deep learning according to claim 4, characterized in that, The PC-side GUI view includes a parameter and task panel, a main display area, and a multi-view stack.

6. The blood smear image segmentation and intelligent interactive diagnostic system based on deep learning according to claim 5, characterized in that, The parameters and task panel includes image selection, confidence threshold, NMS and IoU, target size and morphological kernel, running detection, saving results, and launching the AI ​​assistant.

7. The blood smear image segmentation and intelligent interactive diagnostic system based on deep learning according to claim 6, characterized in that, The main display area includes a detection result view, a toolbar, a detection box, and detection information; The detection results view displays the overall detection results for malaria parasites. It can overlay malaria parasite masks and detection boxes. At the same time, the main display area displays a summary of detection information and provides patient-readable instructions. The toolbar supports zooming, side-by-side comparison, and export. The detection box is used to outline the area containing malaria parasites when detecting a target, and is a magnified view of a local area in the detection results view.

8. The blood smear image segmentation and intelligent interactive diagnostic system based on deep learning according to claim 7, characterized in that, The multi-view stack includes a white blood cell gallery, a parasite candidate queue, and a mask and purification image.

9. The blood smear image segmentation and intelligent interactive diagnostic system based on deep learning according to claim 8, characterized in that, The mask and purification diagram includes a parasite mask 1 and a leukocyte removal diagram.

10. A blood smear image segmentation and intelligent interactive diagnostic system based on deep learning according to claim 9, characterized in that, The linkage rules of the PC-side GUI include: When hovering over / clicking the parasite candidate queue, the overall detection results view is simultaneously positioned and highlighted; The mask and cleaned image support transparency sliders and side-by-side comparison of the original and cleaned images; In the PC GUI, all views share coordinates and zoom and translate synchronously. PC GUI Parameters - Result Snapshot Replay: Each parameter change automatically generates a parameter snapshot, supports one-click rollback and pairwise comparison, and writes it to the case audit log; The review of candidate priorities for the PC GUI includes: Overall score: Score = α·Conf + β·Ctx Where Conf represents the model confidence score; Ctx represents the context score. Candidate priority review supports batch approval, rejection, and re-examination. Approved items are entered into the evidence panel and used for reporting. Concentration estimation and statistics for PC-based GUI, including: The main display area shows the estimated concentration of malaria parasites, positive counts, and confidence intervals. By default, the calculation uses the number of positive candidates per unit field of view / number of reference cells or area normalization. The coefficients are configurable. The AI ​​assistant for the PC GUI and disclaimers include: Summarize current evidence and precautions to generate a "DetectionInformationSummary", clearly stating that it is "for reference only and cannot replace diagnosis by medical personnel".

Images