A method for malaria parasite typing identification based on multi-scale feature extraction and confidence weighting

Abstract

The application discloses a method for typing and identifying malaria parasites based on multi-scale feature extraction and confidence weighting, which combines the advantages of thick blood film and thin blood film dual-mode recognition on the basis of the Ultralytics YOLOv11 model. First, the malaria parasite characteristic digital enhancement modulation strengthens the features, and multi-scale feature extraction and result cleaning are respectively performed on thick blood films (adapted to 9 types of labels) and thin blood films (adapted to 18 types of labels). Then, based on the confidence and quantity weighting rules, preliminary determination of thick blood film positive and negative and preliminary determination of thin blood film typing are completed. Finally, through cross-blood film feature fusion and comprehensive weighted determination, positive and negative and typing confirmation are realized, and artificial experience digital rules are integrated. The method solves the limitations of single blood film recognition, takes into account the screening sensitivity and identification accuracy, and improves the efficiency and accuracy of malaria parasite recognition and typing.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, specifically to a method for identifying Plasmodium genotyping based on multi-scale feature extraction and confidence weighting. Background Technology

[0002] Malaria parasite detection is a crucial step in malaria prevention and treatment. While blood membrane microscopy is the traditional gold standard, its efficiency and accuracy heavily rely on the experience of laboratory personnel, making it unsuitable for large-scale screening and grassroots medical settings. With the development of artificial intelligence, computer vision-based intelligent malaria parasite identification solutions have gradually become a research hotspot. Among these, the Ultralytics YOLO series models are widely used in medical image analysis due to their high-efficiency target detection capabilities. Existing malaria parasite identification technologies mainly suffer from the following problems: First, insufficient adaptability to blood membrane image features. In thick blood membranes, overlapping cells and hemolysis leave only outlines, resulting in blurred morphological features of the malaria parasite. While thin blood membranes can clearly present the parasite structure, most existing technologies only process thin blood membranes, failing to fully utilize the high sensitivity of thick blood membranes and wasting dual-modal data resources. Second, limited accuracy of feature extraction. Traditional methods lack dedicated extraction modules for image differences between thick and thin blood films, resulting in insufficient multi-scale feature capture capabilities. This makes it difficult to effectively distinguish different developmental stages of Plasmodium (ring bodies, trophozoites, schizonts, gametophytes, etc.) and their typing characteristics, and they are easily affected by impurities and red blood cell fragments. Thirdly, the positive / negative determination and typing logic are simplistic. Existing technologies lack a scientific confidence-weighted mechanism, relying heavily on the identification results of a single blood film. The lack of a comprehensive determination system for thick and thin blood films leads to inaccurate calculations of positive probabilities and potential misjudgments or inability to provide results in typing. Fourthly, the conversion of human experience is insufficient. The professional judgment experience of laboratory personnel has not been effectively digitized and integrated into the identification process. There is a lack of dynamic correction mechanisms for identification results, and no auxiliary technologies have been introduced to improve feature discrimination, further limiting the improvement of identification efficiency. These problems directly result in shortcomings in screening sensitivity, typing accuracy, and clinical applicability of existing technologies, making it difficult to meet actual diagnostic and treatment needs. Summary of the Invention

[0003] The purpose of this invention is to provide a method for identifying Plasmodium genotyping based on multi-scale feature extraction and confidence weighting, so as to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: A method for identifying Plasmodium genotyping based on multi-scale feature extraction and confidence weighting, wherein the identification method is as follows: Step 1: Blood membrane sample preprocessing and digital enhancement modulation of malaria parasite features Thick and thin blood film samples were collected and standardized for slide preparation and staining to ensure image clarity and consistency. Based on computer vision and digital image processing technology, malaria parasite feature digital enhancement modulation was performed on the preprocessed blood film images: the image color, grayscale, and texture channels were split using a multi-channel pixel analysis algorithm. Adaptive contrast gain adjustment, edge feature enhancement filtering, and target region pixel weight allocation strategies were adopted to amplify the digital feature differences between malaria parasites and red blood cells, improving the image recognition of target structures and laying a precise data foundation for subsequent feature extraction. Simultaneously, a basic detection framework was built based on the Ultralytics YOLOv11 model, and model initialization and parameter presets (including feature extraction thresholds, target detection anchor points, and other configurations adapted to the blood film images) were completed. Step 2: Multi-scale feature extraction and label adaptation of thick blood films To address the image characteristics of thick-film cell accumulation and interwoven contours, a dedicated multi-scale feature extraction module was designed. By combining convolutional kernels of different receptive fields, it captures multi-scale morphological features of Plasmodium from tiny ring bodies to larger gametophytes. Based on the developmental stage and type of Plasmodium thick-film, nine tags are matched, including ring bodies, amoebic trophozoites, other trophozoites, high-schizonts, low-schizonts, falciparum gametophytes, other gametophytes, neutrophils, and lymphocytes, to achieve precise matching between features and tags. Step 3: Cleaning and preliminary determination of thick blood film identification results The features extracted from thick blood films are used for target detection and annotation to locate suspected Plasmodium areas. Area threshold ranges for each parasite type are set. The annotation results are cleaned to remove impurities that are too small and misidentified areas that are too large. A weighted judgment rule is established based on the confidence level of Plasmodium and the minimum number of samples that meet the confidence level. The positive probability is calculated to complete the preliminary judgment of positive and negative results. If no Plasmodium that meets the conditions is detected after scanning more than the preset number of images, it is directly judged as a high probability negative, and the cleaned annotation image and the original identification data are saved. Step 4: Calculation of Red Blood Cell Area Based on Thin Blood Film Given the characteristic of thin blood film and intact morphology of red blood cells, red blood cells are first identified and labeled. Red blood cells are segmented from the labeled rectangle and the area of ​​a single red blood cell is calculated. Outliers in the red blood cell area data are removed. When the number of identified red blood cells reaches a preset threshold, identification is stopped and the average red blood cell area is calculated, which is used as the benchmark for subsequent screening of malaria parasite area. Step 5: Multi-dimensional feature extraction and double washing of thin blood membrane Employing a multi-dimensional feature extraction strategy adapted to thin blood films, this method simultaneously captures the morphology, texture, color, and spatial distribution features of Plasmodium. It adapts to various sub-labels, including single-nuclear circular bodies, binuclear circular bodies, and edge-attached circular bodies. First, it sets a threshold for the area of ​​the Plasmodium rectangular frame based on the average area of ​​red blood cells for the first area screening. Then, it segments the Plasmodium region to calculate the actual area and performs a second cleaning based on the precise area threshold, further improving the accuracy of the identification results. Step 6: Preliminary identification of thin-film malaria parasites and adaptation to scanning mode Based on the cleaned Plasmodium data and combined with the typing logic provided by experts from the CDC, the Plasmodium typing and corresponding probability were initially determined. Two scanning methods were set: a fast mode and a standard mode. In the fast mode, if the number of Plasmodium images identified and the confidence level of a certain typing reached the preset condition, the scanning was terminated early. In the standard mode, all thin blood membrane images were scanned in sequence, and the results were output in descending order of typing probability. Step 7: Verification of Plasmodium Feature Digital Enhancement Modulation and Fusion of Transmembrane Features Leveraging the advantages of data enhanced and modulated by digital features of Plasmodium, cross-validation was performed on the digital features of Plasmodium identified in thick and thin blood films (including morphological contour pixel matrix, texture feature vector, and color distribution parameters). By calculating the cosine similarity of feature vectors, the consistency of digital features of suspected Plasmodium targets in the two types of blood films was compared, and abnormal targets with similarity below a preset threshold were eliminated to ensure feature reliability. A cross-blood film feature fusion module was constructed, and an attention-weighted fusion algorithm was adopted. Based on the recognition confidence of the high-sensitivity screening features of thick blood films and the high-resolution typing features of thin blood films, the fusion weight of digital features was dynamically allocated. The effective feature vectors of the two types of blood films were spliced ​​and weighted to form a comprehensive feature vector with both screening breadth and typing accuracy, realizing the complementarity of dual-modal digital features. Step 8: Confidence-weighted comprehensive judgment and classification confirmation Based on the preliminary positive / negative determination results of thick-film malaria parasites, a comprehensive weighted formula is established by combining the number of parasites in thin-film malaria parasites, confidence level, and typing information to correct the positive probability and complete the final positive / negative determination. If the number of thin-film malaria parasite samples meets the typing criteria, all possible typings and probabilities are output; if the number of samples is insufficient and only circular parasites are detected, it is judged to be highly probable as falciparum malaria; in other cases, it is marked that no typing result can be given, and human experience-based digital rules are introduced to fine-tune the determination results. Step 9: Results Integration and Dynamic Optimization Display By integrating the identification data, judgment process, and final results of thick and thin blood films, a comprehensive report is generated, including the probability of positive and negative results, the typing results, and the feature visualization. A result feedback mechanism is established to input the actual clinical diagnosis results back into the model, dynamically optimize the feature extraction parameters, confidence thresholds, and judgment rules, and continuously improve the model's recognition accuracy and robustness.

[0005] Based on the above, further explanation is provided: Malaria parasite feature digital enhancement modulation is based on a purely digital processing flow, using computer algorithms to enhance features and avoid the risk of sample damage. The modulation process includes three core stages: The first stage is channel splitting and feature parsing, splitting the RGB and HSV channels of the image and extracting the grayscale histogram and texture matrix of the malaria parasite target; the second stage is dynamic parameter adjustment, focusing on contour feature enhancement for thick blood film images, using a combination of low-pass filtering and edge sharpening algorithms to highlight the pixel differences in the malaria parasite contour in cell-accumulated scenes; for thin blood film images, focusing on internal structural feature enhancement, using high-pass filtering and local contrast enhancement algorithms to clearly present the pixel distribution of fine intracellular structures; the third stage is feature fusion output, re-integrating the multi-channel enhanced feature data to generate an enhanced standardized blood film image. During model initialization, pre-training is performed on a malaria parasite medical image dataset, with preset learning rate, iteration count, and feature extraction anchor point parameters to ensure that the Ultralytics YOLOv11 model quickly adapts to the detection scenario of the enhanced blood film image, improving target recognition response speed and feature matching accuracy.

[0006] Based on the above, further explanation is provided: the multi-scale feature extraction module contains three parallel convolutional branches, which use 1×1, 3×3, and 5×5 convolutional kernels respectively. Features at different scales are concatenated through a feature fusion layer. During the label adaptation process, an attention mechanism is introduced to give higher weights to the features corresponding to the high-discrimination labels of Plasmodium, thereby increasing the priority of key feature extraction and reducing interference from non-target labels.

[0007] Based on the above, further explanation is provided: the area threshold range was determined by statistically analyzing the area distribution of a large number of normal Plasmodium samples, and extreme values ​​were eliminated using the interquartile range method to ensure the scientific validity of the threshold; when calculating the positive probability, a higher weight coefficient was assigned to high-confidence Plasmodium, and the weight was positively correlated with the confidence level, thereby increasing the influence of high-confidence targets on the judgment results; the preset number of scans was set based on the routine testing volume of thick blood film samples.

[0008] Based on the above, further explanation is provided: Red blood cell segmentation employs an adaptive threshold segmentation algorithm, combined with edge detection technology to improve segmentation accuracy; outlier removal uses the Z-score method, identifying red blood cells as outliers when their area exceeds the mean plus or minus three times the standard deviation; and the preset red blood cell count threshold is determined through statistical analysis to ensure a balance between reliability and efficiency in average area calculation.

[0009] Based on the above, further explanation is provided: In multi-dimensional feature extraction, texture features are calculated using a gray-level co-occurrence matrix, color features are extracted based on the HSV color space, and spatial distribution features are achieved through coordinate encoding; the threshold setting for double cleaning adopts a dynamic adjustment mechanism, which is adaptively optimized according to the quality differences of different batches of blood membrane samples to avoid screening bias caused by a fixed threshold.

[0010] Based on the above, further explanation is provided: During the initial classification determination, feature correlation analysis is introduced to calculate the matching degree between different classification features and detection features, and the weighted results are incorporated into the determination; the preset conditions of the fast mode are determined through verification with a large number of clinical samples to ensure that the accuracy of the determination is not reduced while improving efficiency; in the standard mode, the classification probabilities of all images are accumulated and averaged to reduce the impact of single image errors on the results.

[0011] Based on the above, further explanation is provided: feature cross-validation uses cosine similarity calculation. When the similarity of the corresponding features of thick and thin blood membranes is lower than a preset threshold, it is marked as a suspected target and re-validated. Cross-blood membrane feature fusion adopts an attention-weighted fusion strategy, which dynamically allocates fusion weights according to the recognition confidence of the two blood membranes. The higher the confidence, the greater the weight of the blood membrane feature.

[0012] Based on the above, further explanation is provided: the comprehensive weighted formula incorporates three core indicators: the positive probability of thick blood film, the confidence level of thin blood film typing, and the similarity of feature fusion. The weight coefficients of each indicator are optimized through the gradient descent algorithm. The digitized rules based on human experience include correction thresholds and judgment priorities for different testing scenarios, which can be personalized according to clinical needs to ensure the practicality of the results.

[0013] Based on the above, further explanation is provided: the results are presented using a combination of visual charts and text descriptions to highlight key features and judgment criteria; during the dynamic optimization process, an incremental learning strategy is adopted, and the model is only fine-tuned for newly added feedback data to avoid repeated training. At the same time, the parameter adjustment trajectory is recorded for easy tracking and rollback, ensuring the stability of model optimization.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This technical solution focuses on the core elements of identifying the combination of thick and thin blood films, digitizing human experience, and operating based on the UltralyticsYOLOv11 model, achieving significant technical results.

[0015] In terms of recognition efficiency, it breaks through the existing recognition mode that mainly relies on thin blood membranes. By complementing the dual modalities of thick and thin blood membranes and combining them with a rapid scanning mode, recognition can be terminated early when preset conditions are met, achieving rapid judgment and significantly shortening the detection time to meet the needs of large-scale screening. At the same time, based on the efficient detection capabilities of the Ultralytics YOLOv11 model, coupled with a dedicated multi-scale feature extraction module, the speed of feature extraction and target recognition is further improved.

[0016] Regarding recognition accuracy, an adapted segmentation and detection module and a dedicated labeling system were designed to address the differences in image features between the two types of blood membranes. Abnormal targets were eliminated through double cleaning, and the false positive rate was effectively reduced by combining confidence-weighted judgment rules. The image interpretation experience of human instructors was transformed into digital rules and correction thresholds, which were integrated into the comprehensive judgment and dynamic optimization process, significantly improving the reliability of positive and negative determinations and typing results, while taking into account the advantages of high-sensitivity screening of thick blood membranes and accurate identification of thin blood membranes.

[0017] Furthermore, this method, through cross-blood membrane feature fusion and dynamic optimization mechanisms, can continuously adjust parameters based on clinical feedback, enhancing the robustness and adaptability of the model. It solves the problems of single blood membrane recognition, insufficient feature extraction, and inadequate experience transformation in existing technologies, providing more efficient and accurate technical support for malaria diagnosis and treatment. Attached Figure Description

[0018] Figure 1 This is a flowchart of the identification method of the present invention. Detailed Implementation

[0019] 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.

[0020] Example

[0021] Please see Figure 1 The present invention provides a technical solution: A method for identifying Plasmodium genotyping based on multi-scale feature extraction and confidence weighting, wherein the identification method is as follows: Step 1: Blood membrane sample preprocessing and digital enhancement modulation of malaria parasite features Thick and thin blood film samples were collected and standardized for slide preparation and staining to ensure image clarity and consistency. Based on computer vision and digital image processing technology, malaria parasite feature digital enhancement modulation was performed on the preprocessed blood film images: the image color, grayscale, and texture channels were split using a multi-channel pixel analysis algorithm. Adaptive contrast gain adjustment, edge feature enhancement filtering, and target region pixel weight allocation strategies were adopted to amplify the digital feature differences between malaria parasites and red blood cells, improving the image recognition of target structures and laying a precise data foundation for subsequent feature extraction. Simultaneously, a basic detection framework was built based on the Ultralytics YOLOv11 model, and model initialization and parameter presets (including feature extraction thresholds, target detection anchor points, and other configurations adapted to the blood film images) were completed. Step 2: Multi-scale feature extraction and label adaptation of thick blood films To address the image characteristics of thick-film cell accumulation and interwoven contours, a dedicated multi-scale feature extraction module was designed. By combining convolutional kernels of different receptive fields, it captures multi-scale morphological features of Plasmodium from tiny ring bodies to larger gametophytes. Based on the developmental stage and type of Plasmodium thick-film, nine tags are matched, including ring bodies, amoebic trophozoites, other trophozoites, high-schizonts, low-schizonts, falciparum gametophytes, other gametophytes, neutrophils, and lymphocytes, to achieve precise matching between features and tags. Step 3: Cleaning and preliminary determination of thick blood film identification results The features extracted from thick blood films are used for target detection and annotation to locate suspected Plasmodium areas. Area threshold ranges for each parasite type are set. The annotation results are cleaned to remove impurities that are too small and misidentified areas that are too large. A weighted judgment rule is established based on the confidence level of Plasmodium and the minimum number of samples that meet the confidence level. The positive probability is calculated to complete the preliminary judgment of positive and negative results. If no Plasmodium that meets the conditions is detected after scanning more than the preset number of images, it is directly judged as a high probability negative, and the cleaned annotation image and the original identification data are saved. Step 4: Calculation of Red Blood Cell Area Based on Thin Blood Film Given the characteristic of thin blood film and intact morphology of red blood cells, red blood cells are first identified and labeled. Red blood cells are segmented from the labeled rectangle and the area of ​​a single red blood cell is calculated. Outliers in the red blood cell area data are removed. When the number of identified red blood cells reaches a preset threshold, identification is stopped and the average red blood cell area is calculated, which is used as the benchmark for subsequent screening of malaria parasite area. Step 5: Multi-dimensional feature extraction and double washing of thin blood membrane Employing a multi-dimensional feature extraction strategy adapted to thin blood films, this method simultaneously captures the morphology, texture, color, and spatial distribution features of Plasmodium. It adapts to various sub-labels, including single-nuclear circular bodies, binuclear circular bodies, and edge-attached circular bodies. First, it sets a threshold for the area of ​​the Plasmodium rectangular frame based on the average area of ​​red blood cells for the first area screening. Then, it segments the Plasmodium region to calculate the actual area and performs a second cleaning based on the precise area threshold, further improving the accuracy of the identification results. Step 6: Preliminary identification of thin-film malaria parasites and adaptation to scanning mode Based on the cleaned Plasmodium data and combined with the typing logic provided by experts from the CDC, the Plasmodium typing and corresponding probability were initially determined. Two scanning methods were set: a fast mode and a standard mode. In the fast mode, if the number of Plasmodium images identified and the confidence level of a certain typing reached the preset condition, the scanning was terminated early. In the standard mode, all thin blood membrane images were scanned in sequence, and the results were output in descending order of typing probability. Step 7: Verification of Plasmodium Feature Digital Enhancement Modulation and Fusion of Transmembrane Features Leveraging the advantages of data enhanced and modulated by digital features of Plasmodium, cross-validation was performed on the digital features of Plasmodium identified in thick and thin blood films (including morphological contour pixel matrix, texture feature vector, and color distribution parameters). By calculating the cosine similarity of feature vectors, the consistency of digital features of suspected Plasmodium targets in the two types of blood films was compared, and abnormal targets with similarity below a preset threshold were eliminated to ensure feature reliability. A cross-blood film feature fusion module was constructed, and an attention-weighted fusion algorithm was adopted. Based on the recognition confidence of the high-sensitivity screening features of thick blood films and the high-resolution typing features of thin blood films, the fusion weight of digital features was dynamically allocated. The effective feature vectors of the two types of blood films were spliced ​​and weighted to form a comprehensive feature vector with both screening breadth and typing accuracy, realizing the complementarity of dual-modal digital features. Step 8: Confidence-weighted comprehensive judgment and classification confirmation Based on the preliminary positive / negative determination results of thick-film malaria parasites, a comprehensive weighted formula is established by combining the number of parasites in thin-film malaria parasites, confidence level, and typing information to correct the positive probability and complete the final positive / negative determination. If the number of thin-film malaria parasite samples meets the typing criteria, all possible typings and probabilities are output; if the number of samples is insufficient and only circular parasites are detected, it is judged to be highly probable as falciparum malaria; in other cases, it is marked that no typing result can be given, and human experience-based digital rules are introduced to fine-tune the determination results. Step 9: Results Integration and Dynamic Optimization Display By integrating the identification data, judgment process, and final results of thick and thin blood films, a comprehensive report is generated, including the probability of positive and negative results, the typing results, and the feature visualization. A result feedback mechanism is established to input the actual clinical diagnosis results back into the model, dynamically optimize the feature extraction parameters, confidence thresholds, and judgment rules, and continuously improve the model's recognition accuracy and robustness.

[0022] Analysis of the above technical content: This solution mainly adopts the Ultralytics YOLOv11 model as the basic detection framework, combined with Plasmodium feature digital enhancement modulation technology, multi-scale and multi-dimensional feature extraction strategies, confidence weighting judgment rules, and cross-blood membrane feature fusion mechanism. This solves the core problems in existing Plasmodium identification technologies, such as low utilization rate of single blood membrane data, insufficient accuracy of feature extraction, simplistic logic for positive / negative and genotyping judgments, and insufficient conversion of human experience. Plasmodium feature digital enhancement modulation technology is applied to enhance the features of blood membrane samples. By dynamically adjusting the signal frequency, the contrast between the contour features of thick blood membranes and the internal structural features of thin blood membranes is specifically improved, laying a crucial foundation for subsequent accurate identification. An innovative dedicated processing module adapted to the differences between thick and thin blood membranes is designed. On the thick blood membrane side, a combination of 1×1, 3×3, and 5×5 multi-scale convolutional kernels and a 9-class label system efficiently captures the multi-scale morphological features of Plasmodium in cell accumulation scenarios. On the thin blood membrane side, red blood cell area benchmark calibration and multi-dimensional feature extraction of morphology, texture, color, and spatial distribution are used, along with 18... The class-specific labeling and dual area cleaning mechanism enable the accurate capture of fine features of Plasmodium. At the same time, for the first time, confidence and quantity weighting rules, feature correlation analysis, and cosine similarity cross-validation are systematically integrated into the dual-modal recognition process. By dynamically allocating fusion weights, the advantages of high-sensitivity screening with thick blood films and high-resolution typing with thin blood films are complemented. Ultimately, the efficiency, accuracy and robustness of Plasmodium positivity determination and typing identification are synergistically improved, effectively meeting the actual needs of large-scale screening and primary healthcare scenarios.

[0023] Firstly, the numerical values ​​of digital enhancement modulation of malaria parasite features are controlled within a safe threshold range to avoid damaging blood membrane samples; the signal frequency is dynamically adjusted according to the blood membrane type, with lower frequencies used to enhance contour features in thick blood membranes and higher frequencies used to enhance internal structural features in thin blood membranes; during model initialization, pre-training is performed based on the malaria parasite medical image dataset, with preset learning rate and iteration number parameters to ensure that the Ultralytics YOLOv11 model can quickly adapt to blood membrane detection scenarios.

[0024] Secondly, the multi-scale feature extraction module contains three parallel convolutional branches, which use 1×1, 3×3 and 5×5 convolutional kernels respectively. The feature fusion layer splices features of different scales. During the label adaptation process, an attention mechanism is introduced to give higher weight to the features corresponding to the high-identity labels of Plasmodium, thereby increasing the priority of key feature extraction and reducing interference from non-target labels.

[0025] Thirdly, the area threshold range is determined by statistically analyzing the area distribution of a large number of normal Plasmodium samples. The interquartile range method is used to eliminate extreme values ​​to ensure the scientific validity of the threshold. When calculating the positive probability, a higher weight coefficient is assigned to high-confidence Plasmodium, and the weight is positively correlated with the confidence level, thereby increasing the influence of high-confidence targets on the judgment results. The preset number of scans is set according to the routine testing volume of thick blood film samples.

[0026] Fourthly, red blood cell segmentation adopts an adaptive threshold segmentation algorithm, combined with edge detection technology to improve segmentation accuracy; outlier removal adopts the Z-score method, and red blood cell area exceeding the mean plus or minus 3 times the standard deviation is judged as an outlier; the preset red blood cell number threshold is determined through statistical analysis to ensure a balance between reliability and efficiency in average area calculation.

[0027] Fifthly, in multi-dimensional feature extraction, texture features are calculated using the gray-level co-occurrence matrix, color features are extracted based on the HSV color space, and spatial distribution features are achieved through coordinate encoding; the threshold setting for double cleaning adopts a dynamic adjustment mechanism, which is adaptively optimized according to the quality differences of different batches of blood membrane samples to avoid screening bias caused by a fixed threshold.

[0028] Sixth: In the initial classification determination, feature correlation analysis is introduced to calculate the matching degree between different classification features and detection features, and the weighted results are incorporated into the determination results; the preset conditions of the fast mode are determined through a large number of clinical samples to ensure that the accuracy of the determination is not reduced while improving efficiency; in the standard mode, the classification probabilities of all images are accumulated and averaged to reduce the impact of single image errors on the results.

[0029] Seventh: Feature cross-validation uses cosine similarity calculation. When the similarity of the corresponding features of thick and thin blood membranes is lower than the preset threshold, it is marked as a suspected target and re-validated. Cross-blood membrane feature fusion adopts attention-weighted fusion strategy, which dynamically allocates fusion weights according to the recognition confidence of the two blood membranes. The higher the confidence, the greater the weight of the blood membrane feature.

[0030] Eighth: The comprehensive weighted formula incorporates three core indicators: the positive probability of thick blood film, the confidence level of thin blood film typing, and the similarity of feature fusion. The weight coefficients of each indicator are optimized through the gradient descent algorithm. The digital rules based on human experience include correction thresholds and judgment priorities for different testing scenarios, which can be personalized according to clinical needs to ensure the practicality of the results.

[0031] Ninth: The results are presented using a combination of visual charts and text descriptions to highlight key features and judgment criteria; during the dynamic optimization process, an incremental learning strategy is adopted, and the model is only fine-tuned for newly added feedback data to avoid repeated training. At the same time, the parameter adjustment trajectory is recorded to facilitate traceability and rollback, ensuring the stability of model optimization.

[0032] Analysis of the above technical content: The core of this solution revolves around the deep fusion of dual-modal data and the optimization of intelligent judgment logic. It adopts a threshold setting method driven by statistical analysis, digital human experience rules and dynamic feedback optimization mechanism, which further solves the derivative problems caused by fixed thresholds in the existing technology, such as screening bias, reliance on human subjective judgment, and insufficient model adaptability. At the data processing level, the innovative approach employs the interquartile range method to determine the malaria parasite area threshold and the Z-score method to remove abnormal red blood cell area values. Combined with a dynamic threshold adjustment mechanism, this ensures the scientific rigor and consistency of processing blood membrane samples from different quality batches. At the judgment logic level, for the first time, the typing experience of experts from the CDC is transformed into quantifiable 10-category markers and 11-dimensional typing judgment formulas. Through confidence accumulation capping, step adjustment, and difference optimization strategies, accurate calibration of typing results is achieved. Simultaneously, the introduction of human experience to digitally correct thresholds and judgment priorities allows for personalized configuration based on clinical scenarios, significantly improving the practicality of the technical solution. At the model iteration level, the innovative approach uses an incremental learning strategy, inputting clinical diagnosis results back into the model to dynamically optimize feature extraction parameters, confidence thresholds, and judgment rules. This avoids the waste of resources from repeated training and ensures optimization stability through parameter adjustment trajectory tracking and rollback mechanisms. Ultimately, this achieves a technological leap from "single detection" to "precise diagnosis + dynamic iteration" in malaria parasite identification, providing an intelligent solution for malaria prevention and treatment that combines efficiency, accuracy, and adaptability.

[0033] I. Working Principle of this Solution This solution uses the Ultralytics YOLOv11 model as its basic detection framework. Its core revolves around the fusion of thick and thin blood membrane data, employing multi-dimensional techniques to achieve accurate identification and typing of malaria parasites. The overall workflow is progressive and interconnected. First, blood membrane samples are preprocessed and digitally enhanced to strengthen malaria parasite features. Thick and thin blood membrane samples are standardized in slide preparation and staining to ensure consistent image quality. Then, the signal frequency is dynamically adjusted according to the blood membrane type—a lower frequency is used to enhance contour features in thick blood membranes, while a higher frequency is used to enhance internal structural features in thin blood membranes. By altering the differences in characteristics between malaria parasites and red blood cells, image feature contrast is improved. Simultaneously, model initialization and parameter preset are completed.

[0034] To address the characteristics of thick blood film cell accumulation and interwoven contours, a dedicated multi-scale feature extraction module was designed. By combining 1×1, 3×3, and 5×5 parallel convolutional kernels, multi-scale morphological features of Plasmodium at different developmental stages are captured, and nine dedicated labels are adapted to achieve accurate matching. Subsequently, target detection and labeling are performed on the extracted features. The results are cleaned according to a preset area threshold to exclude abnormal areas. Then, a weighted rule is established based on the confidence level of Plasmodium and the minimum number of samples that meet the confidence level to complete the preliminary positive and negative determination. If no target is detected in consecutive scans exceeding a preset number, it is directly determined as a high-probability negative.

[0035] For thin blood membranes, red blood cells are first identified and segmented using an adaptive threshold segmentation algorithm. The Z-score method is used to remove outliers in area. When the number of red blood cells meets the standard, the average area is calculated as the screening benchmark. Then, a multi-dimensional feature extraction strategy is used to simultaneously capture the morphological, texture, color, and spatial distribution features of the malaria parasite, adapting to 18 subcategories of labels. The results are cleaned through two area screenings (rectangular screening based on the average area of ​​red blood cells and precise screening based on actual area). Finally, the classification and probability are preliminarily determined by combining the classification logic of experts from the CDC, and two scanning modes are provided: fast and standard. In fast mode, the scan can be terminated early if preset conditions are met, while in standard mode, the full scan is completed and the results are output in probability order.

[0036] The process then proceeds to the cross-blood membrane fusion and comprehensive judgment stage. Utilizing the enhanced feature differences resulting from digitally augmented modulation of Plasmodium characteristics, cosine similarity calculations are used to cross-validate the thick and thin blood membrane identification features, eliminating inconsistent suspected targets. An attention-weighted fusion strategy is employed to construct a cross-blood membrane feature fusion module, dynamically assigning weights based on the confidence levels of the two blood membrane identification methods, achieving complementarity between high-sensitivity screening features and high-resolution typing features. Finally, a comprehensive weighted formula is established, incorporating three major indicators: the positive probability of thick blood membrane, the typing confidence level of thin blood membrane, and feature fusion similarity. The results are then fine-tuned using digital rules based on human experience to complete the final positive / negative and typing determination. Simultaneously, a result feedback mechanism is established, inputting clinical diagnosis results back into the model, and using an incremental learning strategy to dynamically optimize parameters and continuously improve model performance.

[0037] II. Core Innovations of this Plan A dual-modal fusion recognition scheme for thick and thin blood membranes is proposed, addressing the issue that most existing technologies only target thin blood membrane recognition and fail to fully utilize the high-sensitivity screening advantages of thick blood membranes. By designing processing modules adapted to the image features of the two blood membranes respectively, thick blood membranes focus on high-sensitivity screening, while thin blood membranes focus on high-precision typing. Then, cross-blood membrane feature fusion is performed to achieve complementary advantages, avoiding the limitations of single blood membrane recognition and maximizing the utilization of dual-modal data resources.

[0038] An innovative multi-scale and multi-dimensional feature extraction strategy was designed to address the shortcomings of traditional methods, such as insufficient accuracy and difficulty in distinguishing different developmental stages and subtypes of Plasmodium. On the thick blood membrane side, morphological features were captured through multi-scale convolutional kernel combinations, while on the thin blood membrane side, multi-dimensional features including morphology, texture, color, and spatial distribution were extracted simultaneously. An attention mechanism was incorporated to prioritize key feature extraction, effectively resisting interference from impurities and red blood cell fragments, significantly improving the comprehensiveness and accuracy of feature extraction.

[0039] A scientifically sound and comprehensive confidence-weighted judgment system has been constructed, solving the problems of simplistic logic in positive / negative determination and typing in existing technologies, as well as inaccurate calculation of positive probability. From the initial confidence-quantity weighted judgment of thick blood films, to the double washing and typing probability calculation of thin blood films, and then to the multi-indicator fusion of the comprehensive weighted formula, a full-process weighted judgment logic is formed. Simultaneously, mechanisms such as feature correlation analysis and dynamic threshold adjustment are introduced to ensure the scientific validity and reliability of the judgment results.

[0040] For the first time, this technology systematically integrates malaria parasite feature enhancement and modulation (DEM) technology with human experience-based digitized rules into the malaria parasite identification process, achieving an innovative breakthrough not addressed in existing technologies. DEM technology enhances feature discrimination by altering biological characteristics, laying the foundation for accurate identification. The human experience-based digitized rules include correction thresholds and judgment priorities for different testing scenarios, which can be customized. Combined with a dynamic optimization mechanism based on incremental learning, it enables continuous iteration of model performance, addressing the industry pain point of insufficient conversion of human experience.

[0041] III. Technical Effects of Implementing this Plan Significantly improved recognition efficiency meets diverse testing needs. Breaking through the traditional single blood membrane recognition mode, it combines rapid screening with flexible scanning with a thick blood membrane. In rapid mode, scanning can be terminated early when the confidence level and number of images are met, greatly shortening the testing time. The high-efficiency detection capability of the Ultralytics YOLOv11 model, combined with a dedicated feature extraction module, further improves the speed of feature extraction and target recognition. At the same time, the preliminary positive and negative determination with a thick blood membrane can quickly filter out a large number of negative samples, reducing the pressure of subsequent thin blood membrane processing. Overall, it is suitable for the efficiency requirements of large-scale screening and primary healthcare scenarios.

[0042] Significant improvements were made in identification accuracy and typing precision. A dedicated processing scheme was designed to address the image differences between the two blood membranes: multi-scale feature extraction and 9-class label adaptation for the thick blood membrane, and multi-dimensional feature extraction and 18-class subdivision label matching for the thin blood membrane, achieving precise capture of Plasmodium features. A dual cleaning mechanism, outlier removal method, and scientifically set area thresholds effectively eliminated impurities and misidentified areas, reducing interference. A confidence-weighted judgment system and a cross-blood membrane feature fusion strategy combined the high sensitivity of the thick blood membrane with the high resolution of the thin blood membrane, significantly reducing the error in positive / negative typing and making the typing results more consistent with reality. The integration of digitized rules based on human experience further corrected judgment biases and improved the reliability of the results.

[0043] The model exhibits strong robustness and clinical adaptability. Through a dynamic threshold adjustment mechanism, the screening criteria can be adaptively optimized based on the quality differences of different batches of blood membrane samples, avoiding screening bias caused by fixed thresholds. The incremental learning strategy enables the model to continuously optimize feature extraction parameters, confidence thresholds, and decision rules based on clinical feedback, adapting to changes in different clinical scenarios without repeated training. Furthermore, the parameter adjustment trajectory is traceable and rollbackable, ensuring optimization stability. The dual-modal fusion design reduces dependence on the quality of a single blood membrane sample; even in cases where some samples are of poor quality, effective judgment can still be achieved by supplementing data from another blood membrane, significantly improving the adaptability and reliability of the technical solution in practical clinical applications.

[0044] This approach standardizes and digitizes human experience, promoting the standardization of malaria parasite detection. It transforms the experience of laboratory personnel in interpreting images and the classification logic of CDC experts into quantifiable judgment rules, correction thresholds, and weighting coefficients. This avoids the inconsistencies caused by the reliance on subjective experience in traditional manual microscopic examination, making the malaria parasite detection process more standardized and repeatable. The generated comprehensive report contains rich information such as positive / negative probabilities, classification results, and feature visualizations, intuitively presenting the judgment criteria and facilitating clinical review and traceability. This provides more scientific and standardized technical support for malaria diagnosis and treatment, effectively addressing the shortage of professional personnel in primary healthcare and large-scale screening.

[0045] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0046] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for identifying Plasmodium genotyping based on multi-scale feature extraction and confidence weighting, characterized in that: The identification method is as follows: Step 1: Blood membrane sample preprocessing and digital enhancement modulation of malaria parasite features Thick and thin blood film samples were collected and standardized for slide preparation and staining to ensure image clarity and consistency. Based on computer vision and digital image processing technology, malaria parasite feature digital enhancement modulation was performed on the preprocessed blood film images: the image color, grayscale, and texture channels were split using a multi-channel pixel analysis algorithm. Adaptive contrast gain adjustment, edge feature enhancement filtering, and target region pixel weight allocation strategies were adopted to amplify the digital feature differences between malaria parasites and red blood cells, improving the image recognition of target structures and laying a precise data foundation for subsequent feature extraction. Simultaneously, a basic detection framework was built based on the Ultralytics YOLOv11 model, and model initialization and parameter presets (including feature extraction thresholds, target detection anchor points, and other configurations adapted to the blood film images) were completed. Step 2: Multi-scale feature extraction and label adaptation of thick blood films To address the image characteristics of thick-film cell accumulation and interwoven contours, a dedicated multi-scale feature extraction module was designed. By combining convolutional kernels of different receptive fields, it captures multi-scale morphological features of Plasmodium from tiny ring bodies to larger gametophytes. Based on the developmental stage and type of Plasmodium thick-film, nine tags are matched, including ring bodies, amoebic trophozoites, other trophozoites, high-schizonts, low-schizonts, falciparum gametophytes, other gametophytes, neutrophils, and lymphocytes, to achieve precise matching between features and tags. Step 3: Cleaning and preliminary determination of thick blood film identification results The features extracted from thick blood films are used for target detection and annotation to locate suspected Plasmodium areas. Area threshold ranges for each parasite type are set. The annotation results are cleaned to remove impurities that are too small and misidentified areas that are too large. A weighted judgment rule is established based on the confidence level of Plasmodium and the minimum number of samples that meet the confidence level. The positive probability is calculated to complete the preliminary judgment of positive and negative results. If no Plasmodium that meets the conditions is detected after scanning more than the preset number of images, it is directly judged as a high probability negative, and the cleaned annotation image and the original identification data are saved. Step 4: Calculation of Red Blood Cell Area Based on Thin Blood Film Given the characteristic of thin blood film and intact morphology of red blood cells, red blood cells are first identified and labeled. Red blood cells are segmented from the labeled rectangle and the area of ​​a single red blood cell is calculated. Outliers in the red blood cell area data are removed. When the number of identified red blood cells reaches a preset threshold, identification is stopped and the average red blood cell area is calculated, which is used as the benchmark for subsequent screening of malaria parasite area. Step 5: Multi-dimensional feature extraction and double washing of thin blood membrane Employing a multi-dimensional feature extraction strategy adapted to thin blood films, this method simultaneously captures the morphology, texture, color, and spatial distribution features of Plasmodium. It adapts to various sub-labels, including single-nuclear circular bodies, binuclear circular bodies, and edge-attached circular bodies. First, it sets a threshold for the area of ​​the Plasmodium rectangular frame based on the average area of ​​red blood cells for the first area screening. Then, it segments the Plasmodium region to calculate the actual area and performs a second cleaning based on the precise area threshold, further improving the accuracy of the identification results. Step 6: Preliminary identification of thin-film malaria parasites and adaptation to scanning mode Based on the cleaned Plasmodium data and combined with the typing logic provided by experts from the CDC, the Plasmodium typing and corresponding probability were initially determined. Two scanning methods were set: a fast mode and a standard mode. In the fast mode, if the number of Plasmodium images identified and the confidence level of a certain typing reached the preset condition, the scanning was terminated early. In the standard mode, all thin blood membrane images were scanned in sequence, and the results were output in descending order of typing probability. Step 7: Verification of Plasmodium Feature Digital Enhancement Modulation and Fusion of Transmembrane Features Leveraging the advantages of data enhanced and modulated by digital features of Plasmodium, cross-validation was performed on the digital features of Plasmodium identified in thick and thin blood films (including morphological contour pixel matrix, texture feature vector, and color distribution parameters). By calculating the cosine similarity of feature vectors, the consistency of digital features of suspected Plasmodium targets in the two types of blood films was compared, and abnormal targets with similarity below a preset threshold were eliminated to ensure feature reliability. A cross-blood film feature fusion module was constructed, and an attention-weighted fusion algorithm was adopted. Based on the recognition confidence of the high-sensitivity screening features of thick blood films and the high-resolution typing features of thin blood films, the fusion weight of digital features was dynamically allocated. The effective feature vectors of the two types of blood films were spliced ​​and weighted to form a comprehensive feature vector with both screening breadth and typing accuracy, realizing the complementarity of dual-modal digital features. Step 8: Confidence-weighted comprehensive judgment and classification confirmation Based on the preliminary positive / negative determination results of thick-film malaria parasites, a comprehensive weighted formula is established by combining the number of parasites in thin-film malaria parasites, confidence level, and typing information to correct the positive probability and complete the final positive / negative determination. If the number of thin-film malaria parasite samples meets the typing criteria, all possible typings and probabilities are output; if the number of samples is insufficient and only circular parasites are detected, it is judged to be highly probable as falciparum malaria; in other cases, it is marked that no typing result can be given, and human experience-based digital rules are introduced to fine-tune the determination results. Step 9: Results Integration and Dynamic Optimization Display By integrating the identification data, judgment process, and final results of thick and thin blood films, a comprehensive report is generated, including the probability of positive and negative results, the typing results, and the feature visualization. A result feedback mechanism is established to input the actual clinical diagnosis results back into the model, dynamically optimize the feature extraction parameters, confidence thresholds, and judgment rules, and continuously improve the model's recognition accuracy and robustness.

2. The method for identifying Plasmodium genotyping based on multi-scale feature extraction and confidence weighting according to claim 1, characterized in that: The digital enhancement modulation of malaria parasite features is based on a purely digital processing flow and uses computer algorithms to enhance features, thus avoiding the risk of sample damage. The modulation process comprises three core stages: the first stage involves channel splitting and feature parsing, splitting the RGB and HSV channels of the image to extract the grayscale histogram and texture matrix of the malaria parasite target; the second stage involves dynamic parameter adjustment, focusing on contour feature enhancement for thick blood film images, employing a combination of low-pass filtering and edge sharpening algorithms to highlight the pixel differences in the malaria parasite contours in cell-accumulated scenes; and focusing on internal structural feature enhancement for thin blood film images, employing high-pass filtering and local contrast enhancement algorithms to clearly present the pixel distribution of fine intracellular structures; the third stage is feature fusion output, which re-integrates the enhanced feature data from multiple channels to generate an enhanced, standardized blood film image. During model initialization, pre-training is performed on a malaria parasite medical image dataset, with preset learning rate, iteration count, and feature extraction anchor point parameters to ensure that the Ultralytics YOLOv11 model can quickly adapt to the detection scenarios of enhanced blood film images, improving target recognition response speed and feature matching accuracy.

3. The method for identifying Plasmodium genotyping based on multi-scale feature extraction and confidence weighting according to claim 1, characterized in that: The multi-scale feature extraction module contains three parallel convolutional branches, which use 1×1, 3×3, and 5×5 convolutional kernels respectively. Features at different scales are concatenated through a feature fusion layer. During the tag adaptation process, an attention mechanism is introduced to assign higher weights to features corresponding to high-identity malaria parasite tags, thereby increasing the priority of key feature extraction and reducing interference from non-target tags.

4. The method for malaria parasite typing and identification based on multi-scale feature extraction and confidence weighting according to claim 1, characterized in that: The area threshold range was determined by statistically analyzing the area distribution of a large number of normal Plasmodium samples, and extreme values ​​were eliminated using the interquartile range method to ensure the scientific validity of the threshold. When calculating the positive probability, a higher weight coefficient is assigned to high-confidence malaria parasites. The weight is positively correlated with the confidence level, which increases the influence of high-confidence targets on the judgment results. The preset number of scans is set according to the routine testing volume of thick blood film samples.

5. The method for identifying Plasmodium genotyping based on multi-scale feature extraction and confidence weighting according to claim 1, characterized in that: Red blood cell segmentation employs an adaptive threshold segmentation algorithm, combined with edge detection technology to improve segmentation accuracy. Outlier removal was performed using the Z-score method, where red blood cell area exceeding the mean plus or minus three standard deviations was considered an outlier. The preset red blood cell count threshold is determined through statistical analysis to ensure a balance between reliability and efficiency in the average area calculation.

6. The method for identifying Plasmodium genotyping based on multi-scale feature extraction and confidence weighting according to claim 1, characterized in that: In multi-dimensional feature extraction, texture features are calculated using the gray-level co-occurrence matrix, color features are extracted based on the HSV color space, and spatial distribution features are achieved through coordinate encoding. The threshold setting for double washing adopts a dynamic adjustment mechanism, which is adaptively optimized according to the quality differences of different batches of blood membrane samples to avoid screening bias caused by a fixed threshold.

7. The method for identifying Plasmodium genotyping based on multi-scale feature extraction and confidence weighting according to claim 1, characterized in that: In the initial classification determination, feature correlation analysis is introduced to calculate the matching degree between different classification features and detection features, and the weighted results are incorporated into the determination results. The preset conditions of the fast mode are determined through a large number of clinical samples to ensure that the accuracy of the determination is not reduced while improving efficiency. In the standard mode, the classification probabilities of all images are accumulated and averaged to reduce the impact of single image errors on the results.

8. The method for identifying Plasmodium genotyping based on multi-scale feature extraction and confidence weighting according to claim 1, characterized in that: Feature cross-validation uses cosine similarity calculation. When the similarity of the features corresponding to thick and thin blood films is lower than a preset threshold, they are marked as suspected targets and re-validated. The cross-blood membrane feature fusion adopts an attention-weighted fusion strategy, which dynamically allocates fusion weights based on the recognition confidence of the two blood membranes. The higher the confidence, the greater the weight of the blood membrane feature.

9. The method for malaria parasite typing and identification based on multi-scale feature extraction and confidence weighting according to claim 1, characterized in that: The comprehensive weighted formula incorporates three core indicators: the positive probability of thick blood film, the confidence level of thin blood film typing, and the similarity of feature fusion. The weight coefficients of each indicator are optimized through the gradient descent algorithm. The digitized rules based on human experience include correction thresholds and judgment priorities for different testing scenarios, which can be personalized according to clinical needs to ensure the practicality of the results.

10. The method for identifying Plasmodium genotyping based on multi-scale feature extraction and confidence weighting according to claim 1, characterized in that: The results are presented using a combination of visual charts and text descriptions to highlight key features and judgment criteria. During the dynamic optimization process, an incremental learning strategy is adopted, and the model is only fine-tuned for newly added feedback data to avoid repeated training. At the same time, the parameter adjustment trajectory is recorded for easy tracking and rollback, ensuring the stability of model optimization.

Images