A machine learning-based immunofixation electrophoresis strip intelligent recognition method

Abstract

This invention belongs to the field of machine learning technology, specifically relating to a machine learning-based intelligent identification method for immunofixation electrophoresis bands. The method uses a single band as the smallest identification unit, training the model with manually labeled normal and abnormal bands to accurately identify the features of each band. Furthermore, the model automatically generates different band patterns based on the identified abnormal bands, achieving automated transformation from single band identification to complete band pattern analysis. This strategy significantly reduces the dependence on large-sample training data, requiring only a relatively small amount of high-quality labeled data to achieve good results. In addition, this invention validated the model in multi-center samples, and by further constructing an online identification system, it achieved intelligent processing of the entire immunofixation electrophoresis process, from pattern recognition to band pattern interpretation.

Description

Technical Field

[0001] This invention belongs to the field of machine learning technology, specifically relating to a machine learning-based intelligent identification method for immunofixation electrophoresis bands. Background Technology

[0002] Immunofixation electrophoresis (IFE) is a crucial method in clinical laboratories for identifying monoclonal immunoglobulins, playing a vital role in the diagnosis of multiple myeloma, primary macroglobulinemia, and other monoclonal proliferative diseases. This technique separates proteins in serum or urine on an electrophoretic medium, then reacts them with specific antibodies against IgG, IgA, IgM, κ, and λ light chains, forming clear precipitate bands, thus determining the presence of monoclonal immunoglobulins (M proteins). The detection of M proteins is not only a diagnostic criterion for plasma cell diseases such as multiple myeloma but also plays a key role in disease subtyping, treatment monitoring, and relapse assessment.

[0003] In traditional laboratory procedures, the interpretation of IFE bands relies primarily on manual visual inspection. Laboratory personnel must observe whether dense, clear, and uniformly positioned concentrated bands appear on the electrophoresis strips to determine the presence of monoclonal components. However, this process is time-consuming and dependent on operator experience; different personnel may have varying interpretations of band concentration, often leading to inconsistent results. This subjectivity is particularly pronounced in samples with complex band patterns, faint bands, or nonspecific background, potentially resulting in false positives or false negatives. This problem is especially evident in multi-center laboratories, necessitating an objective, stable, and widely applicable intelligent identification method.

[0004] In recent years, machine learning has developed rapidly in the fields of medical image and laboratory data analysis. Algorithms based on convolutional neural networks and deep learning have demonstrated superior performance compared to human methods in image feature extraction and morphological recognition, and have achieved significant results in areas such as blood smear cell classification and intelligent diagnosis of pathological images. Although there is currently a limited amount of research on machine learning for immunofixation electrophoresis patterns, systematic model construction, strip-level labeled data training, and multi-center validation studies are still lacking. Summary of the Invention

[0005] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly describe some preferred embodiments.

[0006] As one aspect of the present invention, the present invention provides a machine learning-based intelligent identification method for immunofixation electrophoresis bands, which includes the following steps:

[0007] (1) Obtain the immunofixation electrophoresis image of the object to be identified;

[0008] (2) The image is input into a pre-trained strip detection and classification model. The model outputs the location bounding box information and classification of each strip in the image. The classification includes normal strips and abnormal strips. The strip detection and classification model is a convolutional neural network model based on the YOLO series architecture.

[0009] (3) Perform optical character recognition on the image to obtain lane markings and their position information, wherein the lane markings include at least IgG, IgA, IgM, κ light chain and λ light chain;

[0010] (4) Based on the position bounding box of the strip obtained in step (2) and the lane identifier position obtained in step (3), match each strip to the corresponding lane;

[0011] (5) Based on the matching results of step (4) and the preset clinical classification rules, output the final band identification results.

[0012] As a preferred embodiment of the method described in this invention, in step (2), the strip detection and classification model is trained through the following steps:

[0013] (2.1) Constructing the training dataset: Label individual bands from the immunofixation electrophoresis pattern, and label the location bounding box and category label for each band, the category label including normal and abnormal; divide the labeled data into training set, validation set and test set;

[0014] (2.2) Model training: The convolutional neural network model is trained using the training set and optimized using a combined loss function, which includes bounding box regression loss, target confidence loss and classification loss;

[0015] (2.3) Model validation and testing: Use the validation set to adjust the model hyperparameters and use the test set to evaluate the model performance.

[0016] As a preferred embodiment of the method described in this invention, in step (2.2), the bounding box regression loss adopts the CIoU loss function.

[0017] As a preferred embodiment of the method described in this invention, in step (4), the matching of each strip to the corresponding lane is specifically achieved by using the maximum horizontal overlap matching strategy, calculating the horizontal overlap width between the bounding box of each strip and each lane region, and assigning the strip to the lane with the maximum overlap.

[0018] As a preferred embodiment of the method described in this invention, in step (5), the preset clinical classification rules include: when no abnormal band is matched in all lanes, it is determined to be a negative band type; when an abnormal band is matched in a specific lane combination, it is determined to be the corresponding monoclonal immunoglobulin band type, wherein the band type includes at least one of IgG κ, IgG λ, IgA κ, IgA λ, IgM κ, IgM λ, κ monoclonal, and λ monoclonal.

[0019] As a preferred embodiment of the method described in this invention, in step (3), when the optical character identifies that the lane marking is missing, a coordinate interpolation algorithm based on the spatial position of adjacent known markings is used to reconstruct the marking position of the missing lane.

[0020] As a preferred embodiment of the method described in this invention, in step (2), the network architecture of the strip detection and classification model includes a feature extraction backbone network, a multi-scale feature fusion neck network, and a decoupled task prediction head.

[0021] As a preferred embodiment of the method described in this invention, the feature extraction backbone network is based on the CSPDarkNet structure, and the multi-scale feature fusion neck network adopts a bidirectional feature pyramid structure for feature recalibration and fusion.

[0022] The present invention also provides the application of the method in the preparation of medical devices or software products for assisting in the diagnosis of plasma cell diseases.

[0023] The plasma cell diseases mentioned include at least one of multiple myeloma, Waldenström macroglobulinemia, monoclonal gammaglobulinosis of unknown significance, and light chain amyloidosis.

[0024] The beneficial effects of this invention are as follows: A machine learning-based intelligent band recognition model for immunofixation electrophoresis is constructed. Unlike traditional overall pattern classification, this invention uses a single band as the smallest recognition unit. The model is trained using manually labeled normal and abnormal bands, enabling it to accurately identify the characteristics of each band. Furthermore, the model automatically generates different band patterns based on the identified abnormal bands, achieving automated transformation from single band recognition to complete band pattern analysis. This strategy significantly reduces the dependence on large-sample training data, requiring only a relatively small amount of high-quality labeled data to achieve good results. In addition, this invention validated the model in multi-center samples. By further constructing an online recognition system, the entire process of immunofixation electrophoresis results, from pattern recognition to band pattern interpretation, is made intelligent.

[0025] The model constructed in this invention demonstrated high accuracy in identifying both normal and abnormal bands on the test set, with good consistency in band pattern recognition. The overall performance of the model reached or exceeded that of a junior laboratory physician, approaching the interpretation ability of a physician with three years of experience. In a multicenter validation of 5466 atlases, the model maintained stable accuracy, sensitivity, and specificity (all > 0.91). The online system can perform real-time band selection and output of band patterns, significantly improving interpretation efficiency and standardization. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below, wherein:

[0027] Figure 1 This is a diagram of the YOLOv12 model architecture.

[0028] Figure 2 Construction and performance of the IFE strip intelligent recognition model.

[0029] Figure 3 Visualize and interpret the recognition results of different band types.

[0030] Figure 4 This study compares the performance of machine learning models with that of laboratory physicians with different years of experience in IFE band pattern recognition.

[0031] Figure 5 This is used to evaluate the generalization performance of machine learning models in multi-center IFE graphs.

[0032] Figure 6 Example diagram of an online immunofixation electrophoresis band intelligent recognition system. Detailed Implementation

[0033] To make the above-mentioned objectives, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to specific examples.

[0034] Example 1:

[0035] This study collected 12,748 individual band data from immunofixation electrophoresis (IFE), all of which were individually labeled from the IFE patterns. The individual band data included 8,956 normal bands and 3,792 abnormal bands. The abnormal bands encompassed monoclonal bands from eight common abnormal patterns (IgG κ, IgG λ, IgA κ, IgA λ, IgM κ, IgM λ, κ monoclonal, and λ monoclonal). All individual bands were randomly divided in an 8:1:1 ratio into a training set (7,134 normal, 3,067 abnormal), a validation set (906 normal, 354 abnormal), and a test set (916 normal, 371 abnormal) to construct and evaluate the individual band recognition model. In addition, to evaluate the model's ability to identify band patterns on the entire IFE map, this invention collected IFE map data from three hospitals. The three central maps were routine output images from clinical testing and were used for subsequent band pattern identification and multi-center generalization validation.

[0036] The labeling of single bands and the entire image band pattern was independently completed by three experts with 5–10 years of IFE interpretation experience. This included the "normal / abnormal" judgment of single bands and the judgment of the location of all abnormal bands and the final band pattern combination in the entire image. The final label was decided by majority vote (≥2 experts agreeing was considered the gold standard). When comparing the AI ​​recognition performance with human level, three laboratory physicians with different years of experience (1 year, 3 years, and 5 years) were invited to independently interpret the band patterns of the entire image, and the expert consensus was used as the gold standard to calculate accuracy, sensitivity, and specificity.

[0037] The single-band model was trained using only data from Hospital A (Nanjing Gulou Hospital). All band images underwent uniform processing before being input into the model: single bands were labeled from the IFE atlas. The model was first built on the training set of Hospital A to identify "normal and abnormal" single bands, and the parameters were optimized on the validation set. After validation on the test set, the model was further expanded to the entire IFE atlas to achieve band detection, band classification, and combined band pattern output. The whole-image band pattern recognition model was validated on data from three hospitals: Hospital A (Nanjing Gulou Hospital): 2657 IFE atlases; Hospital B (Huai'an First People's Hospital): 795 cases; Hospital C (Shanghai Pudong Hospital): 2014 cases. Table 1 lists the specific band pattern information of the atlases from the three hospitals. The model used did not require retraining and was directly applied to the abnormal band detection and band category recognition of the atlases from the three centers. Subsequently, the accuracy, sensitivity, and specificity of each band pattern were calculated in the three centers to evaluate the transferability and generalization ability of the model under different instruments and sample sources.

[0038] Table 1

[0039]

[0040] Example 2:

[0041] YOLOv12-based strip detection and classification:

[0042] Input and Preprocessing: During training, input electrophoresis images are uniformly adjusted to 640×640 pixels. To avoid image distortion, letterboxing (preserving the original aspect ratio and performing grayscale filling) is recommended, followed by normalization. Hyperparameter Settings: The SGD optimizer is used, with an initial learning rate of 1e-2 and 5 warmups, a weight decay of 5e-4, and a batch size of 16. The total number of training epochs is 300. Loss Function Weights: The bounding box regression loss weight (λ_box) is set to 0.05, the target confidence loss weight (λ_obj) is 1.0, and the classification loss weight (λ_cls) is 0.5. Post-Inference Processing: During prediction, predicted boxes with a confidence level below 0.25 are filtered out, and non-maximum suppression (NMS) is used to remove duplicate boxes. The IoU threshold for NMS is set to 0.45.

[0043] This invention utilizes the YOLOv12 model to accurately locate and classify protein bands in images. Based on this, optical character recognition (OCR) technology is combined to analyze lane markings, ultimately generating a classification report necessary for clinical diagnosis. The entire processing flow achieves end-to-end automated analysis from raw image input to structured report generation. This module aims to accurately locate protein bands in images and determine whether they are "normal" or "abnormal." Employing the YOLOv12 model as the detector offers the advantage of a single-stage, end-to-end detection process, balancing speed and accuracy. This invention achieves immunofixation electrophoresis image analysis through a refined training strategy.

[0044] Network architecture and feature extraction: such as Figure 1 As shown, the overall architecture of YOLOv12 consists of three core parts: a feature extraction backbone network, a multi-scale feature fusion neck network, and a task-decoupled prediction head. These three work together to efficiently map from the original image input to the detection result output.

[0045] The backbone network is based on the CSPDarkNet topology and employs cross-stage partial connections to optimize gradient flow and reduce computational complexity. The network contains five downsampling stages, each consisting of a convolutional layer, a batch normalization layer, and a SiLU activation function. Given an input tensor... The multi-scale feature pyramid output by the backbone network can be represented as:

[0046]

[0047]

[0048] Among them, the highest spatial resolution The feature map retains rich texture and edge information, which is crucial for detecting weak positives, low-contrast bands, and achieving pixel-level precise localization; while the largest receptive field Feature maps contain richer global semantic information, which helps to accurately classify the pathological nature (normal / abnormal) of bands.

[0049] The neck network employs a bidirectional feature pyramid structure, achieving multi-scale information fusion through feature recalibration. This module contains two parallel paths: a top-down path that transmits semantic information through bilinear interpolation upsampling, and a bottom-up path that preserves spatial details through stride convolution. The feature fusion process is formally represented as follows:

[0050]

[0051] in This indicates a channel splicing operation. For batch normalization layer, This is a 2x upsampling operation. This design ensures that features at each scale are normalized before fusion, significantly improving training stability and feature consistency.

[0052] The detection head employs a fully decoupled design, separating bounding box regression, object confidence prediction, and category classification tasks, effectively mitigating optimization conflicts in multi-task learning. For each fused feature map... The network parallel computing has three independent outputs:

[0053]

[0054]

[0055]

[0056] in, Indicates the bounding box regression branch, Indicates the target prediction branch, Indicates a category branch, and It is a standalone, lightweight convolutional module. The activation function is sigmoid. This task-specific design allows each branch to focus on learning its own optimal feature representation.

[0057] The entire network achieves end-to-end computation via forward propagation:

[0058]

[0059] in For network parameters, This is the set of detection results to be output. This architecture provides a robust feature representation foundation for accurate strip detection through hierarchical feature learning and multi-scale information fusion.

[0060] Loss function: Defines the training dataset ,in This represents the input immunofixation electrophoresis image. This indicates the location annotation information (bounding box coordinates) of all stripes in the image. This indicates the category label corresponding to each band. This represents the number of training samples. (Model) The learning objective is to optimize the loss function. To learn model parameters :

[0061]

[0062] in and This is a hyperparameter used to balance the contributions of different loss terms. It is the bounding box regression loss, which uses the CIoU loss function to optimize the positional accuracy between the predicted box and the ground truth box; It is a target-oriented loss used to assess the confidence that each predicted box contains the target; It is a classification loss, using binary cross-entropy loss, to distinguish between normal and abnormal bands.

[0063] Example 3:

[0064] OCR-based lane assignment and classification decision:

[0065] After the YOLOv12 model completes the localization and preliminary classification of the bands, the core task of this stage is to transform the detected biological signals into diagnostic information with clear clinical significance. This process establishes the correspondence between the bands and lane markers through optical character recognition technology and spatial matching algorithms, and completes the final diagnostic classification based on a clinical knowledge base.

[0066] Swimlane Marker Recognition and Coordinate Reconstruction: The PaddleOCR engine is used for text detection and recognition of the input image. First, the text detection module locates all text regions in the image and generates a set of candidate text boxes. Then, the recognition network performs sequence recognition on the characters in each text box, accurately extracting the swimlane marker letters, which correspond to five immunoglobulin types (IgG, IgA, IgM, κ light chain, and λ light chain).

[0067] To address the issue of missing swimlane markers in real-world images, this invention designs a coordinate reconstruction algorithm based on relative positional relationships. By analyzing the spatial distribution of known swimlane markers, the average distance between adjacent markers is calculated to estimate the standard swimlane width. For missing swimlane markers, coordinate interpolation reconstruction is performed based on the spatial positions of their neighboring known markers to ensure complete swimlane position information is obtained. During coordinate reconstruction, the expected number of swimlanes is 5. If the number of identified swimlanes is less than 2, the positions of all swimlanes _1, _2, ..., _5 are estimated using the statistical average of historical data. If the number of identified swimlanes is greater than or equal to 2, the missing marker is determined by the spacing between the identified swimlanes, and then the x-coordinate x_k = x_{k-1} + d is estimated based on the adjacent spacing d.

[0068] Strip-Lane Matching Strategy: Based on the lane spatial information obtained from OCR recognition, this module uses the maximum overlap matching principle to assign the strips detected by YOLOv12 to the corresponding lanes. Specifically, the calculation method involves calculating the horizontal overlap width between the bounding box of each strip and each lane region, i.e., the number of overlapping pixels between the bounding box region and the vertical region of the lane marker. This bounding box is then assigned to the lane with the largest number of overlapping pixels. If there are no overlapping pixels, the horizontal displacement between the vertical region of the lane marker and the bounding box region is considered, and the bounding box is assigned according to the principle of minimum displacement. This calculation method effectively handles problems such as strip position offset and boundary blurring.

[0069] Clinical classification decision: Based on the band-lane matching results, the final classification is performed according to the clinical diagnostic criteria of immunofixation electrophoresis. A negative sample is determined when no abnormal bands are detected in any lane. If abnormal bands are found in one or more lanes, a typing diagnosis is performed based on the corresponding immunoglobulin type, including but not limited to common monoclonal immunoglobulin types such as IgGκ, IgGλ, and IgAκ. For complex abnormal patterns, this method provides suggested diagnostic opinions for clinicians to refer to and confirm.

[0070] Example 4:

[0071] Experimental results of Examples 1-3:

[0072] Construction and performance evaluation of the stripe recognition model:

[0073] 12,748 immunofixation electrophoresis bands were manually labeled, including 8,956 normal bands and 3,792 abnormal bands. The data was divided into training, validation, and test sets in an 8:1:1 ratio. Figure 2A): The training set contains 7,134 normal bands and 3,067 abnormal bands; the validation set contains 906 normal bands and 354 abnormal bands; the test set contains 916 normal bands and 371 abnormal bands. A machine learning model is built based on the training set to distinguish between normal and abnormal bands, and its performance is evaluated on the validation set.

[0074] Figure 2 The following is a description of the construction and performance of the IFE strip intelligent recognition model, including (A) a schematic diagram of the dataset composition, (B) a confusion matrix diagram of the test set, (C) a schematic diagram of the automatic recognition process when the model is extended to the entire spectrum, and (D) a confusion matrix of strip recognition. It can be seen that the model performs well on the test set, and the confusion matrix shows that the recognition accuracy of both normal and abnormal stripes is high. Figure 2 B). Further extending the model to the automatic identification of the entire immunofixation electrophoresis pattern ( Figure 2 C). The model can automatically detect and label normal and abnormal bands in the spectrum, and generate band patterns based on different combinations of abnormal bands. It can identify 8 abnormal band patterns and 1 normal band pattern. A corresponding confusion matrix is ​​generated based on the identification results. Figure 2 (D) The results show that the consistency of the strip pattern recognition is good.

[0075] Visualization and interpretability of model recognition results: To further verify the model's recognition ability and interpretability, the model's recognition results on different band patterns were visualized. Figure 3This invention provides visualization and interpretability of model recognition results for different band patterns. (A–J) showcases the recognition results for 10 typical band patterns, including normal bands and 9 common abnormal band patterns (IgG κ, IgG λ, IgA κ, IgA λ, IgM κ, IgM λ, κ monoclonal, λ monoclonal, and complex pattern). The invention presents recognition results for 10 typical band patterns, including normal bands and 9 common abnormal band types: IgG κ, IgG λ, IgA κ, IgA λ, IgM κ, IgM λ, κ monoclonal, λ monoclonal, and complex pattern. In the visualization results, the model can directly outline the location and extent of abnormal bands on the image, accurately marking dense band regions and their boundaries. Whether it's light chain bands or complex mixed bands, the model can stably identify them and maintain a high degree of consistency with human interpretation results. This indicates that the model not only possesses excellent band detection capabilities but also effectively extracts key morphological features that determine the diagnosis. Furthermore, the results from different band types demonstrate that the model maintains high localization accuracy even in complex scenarios such as weak bands, wide bands, and high background noise, showcasing strong robustness and clinical applicability. This bounding box-based visualization method intuitively reflects the model's discriminative criteria, making the model's output more transparent and interpretable.

[0076] Comparative analysis of model and human interpretation capabilities: To comprehensively evaluate the performance of the machine learning model in actual clinical interpretation, we compared its recognition results with the human recognition results of laboratory physicians with different years of experience (1 year, 3 years, and 10+ years). Figure 4 To compare the performance of the machine learning model with that of laboratory physicians with different experience levels in IFE band pattern recognition, box plots (AC) are used to illustrate the accuracy, sensitivity, and specificity of the model (ML) and three laboratory physicians (N1, N3, N10) across nine band patterns. The comparison covers nine band patterns, including normal bands, IgG κ, IgGλ, IgA κ, IgAλ, IgM κ, IgMλ, κ monoclonal, and λ monoclonal. Both model and human recognition results were calculated using expert consensus as the gold standard, with accuracy, sensitivity, and specificity calculated as shown in Figures 3A–C. Overall, the machine learning model demonstrated stable recognition performance across most band patterns, exhibiting good accuracy, sensitivity, and specificity. Its overall accuracy, sensitivity, and specificity were higher than those of a 1-year experienced laboratory physician and comparable to those of a 3-year experienced physician. This indicates that the model's recognition performance across most band patterns has reached or even surpassed that of junior laboratory physicians and is approaching the interpretation capabilities of intermediate-level physicians.

[0077] Validation of the model’s generalization performance in external multicenters: To verify the generalization performance of the model under different experimental conditions and equipment platforms, this invention further validated it independently in multicenter samples. Figure 5 This section evaluates the generalization performance of the model in multicenter IFE maps. (A–C) shows the accuracy, sensitivity, and specificity of the model in identifying 9 band patterns in hospitals A (n=2657), B (n=795), and C (n=2014). Each graph in the figure represents one band pattern in the map.

[0078] External validation data included 795 immunofixation electrophoresis patterns from Hospital B and 2014 patterns from Hospital C to assess the model's stability and transferability across different centers. During validation, the model identified nine band patterns (Normal, IgG κ, IgG λ, IgA κ, IgA λ, IgM κ, IgM λ, κ monoclonal, and λ monoclonal), and the accuracy, sensitivity, and specificity of each band pattern were calculated. Results showed that the model's overall recognition performance was stable across different centers. Compared to Hospital A, the recognition performance of patterns from Hospitals B and C was slightly lower, but still remained at a high level overall. Hospitals B and C had relatively lower accuracy in identifying normal bands and κ monoclonal bands, but both were above 0.94. Regarding sensitivity, Hospital B's sensitivity for identifying normal bands and κ monoclonal bands was lower than Hospital A, but still above 0.91. In terms of specificity, all three hospitals showed the lowest specificity for identifying κ monoclonal bands, with Hospital B having a specificity of 0.96.

[0079] Online deployment of the model: Figure 6 The following is an example diagram of an online immunofixation electrophoresis band recognition system, where (A) shows an example of the system recognizing a typical normal band, and (B) shows an example of the system recognizing a typical abnormal IgG κ band.

[0080] Building upon the successful model validation, this invention enables the online deployment of an intelligent identification model for immunofixation electrophoresis bands, constructing an intelligent identification system directly applicable to clinical work. This system integrates the model algorithm backend with a web-based interactive interface. Laboratory personnel only need to upload the original immunofixation electrophoresis images, and the system automatically performs image preprocessing, band region localization, feature recognition, and band pattern interpretation, displaying the identification visualization results in real-time in the right-hand window. The system marks the band regions identified by the model with green or red boxes, corresponding to normal and abnormal bands, respectively. Figure 6 A and Figure 6B illustrates the identification of a typical normal band and a typical IgG κ band. The system can accurately identify normal bands and abnormally dense bands, and the identification results are displayed simultaneously in text form, facilitating rapid confirmation by laboratory personnel. In summary, this online identification system automates, visualizes, and instantly outputs immunofixation electrophoresis bands, significantly improving the standardization and ease of use of band interpretation, and providing a feasible technical platform for clinical applications.

[0081] In summary, this invention constructs a machine learning-based intelligent band recognition model for immunofixation electrophoresis (IFE) bands and automates the process from single band identification to inference of band patterns across the entire spectrum. Results show that the model performs stably in band recognition, band pattern classification, and multi-center data validation, significantly improving the objectivity and consistency of IFE interpretation. Traditional IFE interpretation heavily relies on the experience of laboratory personnel, especially when bands are faint, backgrounds are complex, or non-specific interference exists, leading to subjective differences in human interpretation. This study, through band-level annotation and training, enables the model to focus more on local features, thereby improving the sensitivity and interpretability of abnormal bands and helping to reduce the impact of inconsistent human interpretation. The model maintains stable output even with weak bands and complex backgrounds, potentially reducing interpretation differences among beginners, improving reporting efficiency, and providing a more standardized auxiliary tool for the early identification of M protein-related diseases.

[0082] Compared to methods that directly classify based on the entire image, another innovation of this invention lies in decomposing the recognition unit into "single strips." This strategy not only reduces the dependence of training on the amount of data but also makes the model output more transparent. Furthermore, strip-level modeling alleviates, to some extent, the problem of severe imbalance in sample sizes for different band types in IFE, making the model more stable in recognizing rare band types. This invention uses a "band selection" approach as a means of demonstrating model interpretability. The model can directly annotate the dense strip regions of interest in the image. This location-based visualization is more intuitive than heatmaps and closer to the daily interpretation methods of laboratory personnel, while avoiding the diffusion or misleading bright spots that may occur in heatmaps when the strip background is complex. The model maintains good performance in multi-center validation, further demonstrating its stability across devices and experimental conditions, providing a reliable basis for clinical application.

[0083] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A machine learning-based intelligent identification method for immunofixation electrophoresis bands, characterized in that, Includes the following steps: (1) Obtain the immunofixation electrophoresis image of the object to be identified; (2) The image is input into a pre-trained strip detection and classification model. The model outputs the location bounding box information and classification of each strip in the image. The classification includes normal strips and abnormal strips. The strip detection and classification model is a convolutional neural network model based on the YOLO series architecture. (3) Perform optical character recognition on the image to obtain lane markings and their position information, wherein the lane markings include at least IgG, IgA, IgM, κ light chain and λ light chain; (4) Based on the position bounding box of the strip obtained in step (2) and the lane identifier position obtained in step (3), match each strip to the corresponding lane; (5) Based on the matching results of step (4) and the preset clinical classification rules, output the final band identification results.

2. The method according to claim 1, characterized in that, In step (2), the strip detection and classification model is trained through the following steps: (2.1) Constructing a training dataset: Label individual bands from the immunofixation electrophoresis pattern, and label the location bounding box and category label for each band, wherein the category label includes normal and abnormal; The labeled data is divided into training set, validation set and test set; (2.2) Model training: The convolutional neural network model is trained using the training set and optimized using a combined loss function, which includes bounding box regression loss, target confidence loss and classification loss; (2.3) Model validation and testing: Use the validation set to adjust the model hyperparameters and use the test set to evaluate the model performance.

3. The method according to claim 2, characterized in that, In step (2.2), the bounding box regression loss adopts the CIoU loss function.

4. The method according to any one of claims 1-3, characterized in that, In step (4), the matching of each strip to the corresponding lane is specifically achieved by using the maximum horizontal overlap matching strategy. The overlap width between the bounding box of each strip and each lane region in the horizontal direction is calculated, and the strip is assigned to the lane with the maximum overlap.

5. The method according to any one of claims 1-3, characterized in that, In step (5), the preset clinical classification rules include: when no abnormal band is matched in all lanes, it is determined to be a negative band type; when an abnormal band is matched in a specific combination of lanes, it is determined to be the corresponding monoclonal immunoglobulin band type, wherein the band type includes at least one of IgG κ, IgG λ, IgAκ, IgA λ, IgM κ, IgM λ, κ monoclonal, and λ monoclonal.

6. The method according to any one of claims 1-3, characterized in that, In step (3), when the optical character identifies that the lane marking is missing, a coordinate interpolation algorithm based on the spatial position of adjacent known markings is used to reconstruct the marking position of the missing lane.

7. The method according to any one of claims 1-3, characterized in that, In step (2), the network architecture of the strip detection and classification model includes a feature extraction backbone network, a multi-scale feature fusion neck network, and a decoupled task prediction head.

8. The method according to claim 7, characterized in that, The feature extraction backbone network is based on the CSPDarkNet structure, and the multi-scale feature fusion neck network adopts a bidirectional feature pyramid structure for feature recalibration and fusion.

9. The application of the method according to claim 1 in the preparation of medical devices or software products for assisting in the diagnosis of plasma cell diseases.

10. The application according to claim 9, characterized in that, The plasma cell diseases include at least one of multiple myeloma, Waldenström macroglobulinemia, monoclonal gammaglobulinosis of unknown significance, and light chain amyloidosis.

Images