Method and device for recognizing promyelocytic leukemia based on morphological image analysis and storage medium

Abstract

The application discloses a promyelocytic leukemia recognition method and device based on morphological image analysis and a storage medium, and relates to the technical field of medical image processing. The device acquires low-power and high-power images of blood smears and divides the images into slices, extracts a white blood cell target in an HSV space in combination with a maximum inter-class variance method, a Gaussian filter and a morphological open operation double channel, extracts multi-scale features through a residual network, takes low-power global features as a query, adaptively weights and aggregates high-power slices through a gate attention, avoids dilution of sparse pathological features by arithmetic averaging, determines a discount coefficient from information entropy to construct a basic probability distribution containing an unknown state, fuses and calculates a conflict factor through a Dempster combination rule, obtains a trust degree and a likelihood, and finally outputs a recognition result, thereby relieving the problem of high false positive caused by dilution of key features and combination of a fixed threshold and logic or.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method, device and storage medium for identifying promyelocytic leukemia based on morphological image analysis. Background Technology

[0002] Acute promyelocytic leukemia (APLE) is a special type of acute myeloid leukemia characterized by the massive proliferation of abnormal promyelocytes in the bone marrow and peripheral blood, often accompanied by coagulation disorders and severe bleeding tendency. It has a rapid onset and high early mortality rate. When promptly identified and treated with targeted intervention, this disease has a high remission rate, making early screening and preliminary diagnosis clinically significant. For a long time, preliminary diagnosis of this disease has relied primarily on microscopic observation of cell morphology, identifying features such as the lobed and folded shape of the nucleus, the density and distribution of azurophilic granules in the cytoplasm, and the bundled arrangement of rod-shaped bodies. This manual microscopic examination method is labor-intensive, time-consuming, and the results largely depend on the observer's experience, resulting in significant inter- and intra-observer variability.

[0003] With the application of digital pathology and convolutional neural networks in medical image analysis, automated blood smear analysis based on full-field digital slices is increasingly used for auxiliary identification. In the prior art, Chinese invention patent CN118014947B discloses a rapid diagnostic system for identifying the morphology of acute promyelocytic leukemia (APML). This system acquires digital images of blood smears under both low-power (e.g., 10x) and high-power (e.g., 100x) objectives. Four 800×800 pixel image slices are randomly segmented from the cell layer region where single cells are distributed. A ResNet50 residual network, incorporating channel attention and spatial attention mechanisms, identifies each image slice separately. In the diagnostic stage, the system calculates the mean probability that each slice belongs to APML, compares this mean with a pre-set judgment threshold (e.g., 0.3), and combines the judgment results from both low-power and high-power scales using a logical OR operation. Based on this, it outputs the identification conclusion for the entire smear. This system provides a basic framework for automated screening in terms of image segmentation size, double-magnification input, and deep feature extraction from the residual network.

[0004] However, the distribution of abnormal promyelocytes in blood smears is often highly sparse and uneven. When the entire smear is treated as a package and the segmented sections as examples, the above processing method essentially faces a multi-example learning problem. Taking the arithmetic mean of the probabilities of each section indiscriminately implies the assumption that all sections contribute equally to the recognition result. When only a few sections of randomly selected high-magnification sections contain typical lesions and the rest are normal cells, the high probabilities output by a few sections will be significantly lowered by a large number of low probabilities during the averaging process. This causes the sparse but crucial pathological information to be submerged by the normal background, thus significantly reducing the sensitivity of recognition. At the same time, treating the macroscopic distribution of cells reflected at the low magnification scale and the microscopic structure of the nucleocytoplasm reflected at the high magnification scale as two isolated paths and averaging them separately severs the contextual relationship that should exist between the two scales, making it difficult to leverage the synergistic effect of multi-scale information.

[0005] During smear preparation, factors such as batch-to-batch staining variations, uneven coating thickness, red blood cell and platelet overlap, and dust artifacts inevitably affect the network output, introducing inherent uncertainty. The decision-making method described above, which uses a fixed single threshold combined with logical OR operations, determines a positive result as long as the average probability of any scale exceeds the threshold. This fails to account for strong negative information from another scale and lacks a mechanism to measure the reliability of outputs from different scales or to coordinate conflicting evidence. When a scale outputs a high false-positive probability due to smear defects, even if another scale gives an extremely low probability of disease, the system will still force a positive conclusion, resulting in a large number of false positives in batch screening and increasing the burden of subsequent manual review.

[0006] It is evident that the dilution of sparse pathological features due to averaging in the underlying feature aggregation stage, and the insufficient noise resistance and high false positive rate due to fixed thresholds and logical OR operations in the decision-making stage, are two interrelated problems that restrict the practical application of this type of automated identification method. These problems urgently need to be addressed from the perspectives of feature aggregation and decision fusion. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing automated identification methods, such as the dilution of sparse pathological features due to the arithmetic average of the probabilities of each slice, and the insufficient noise resistance and high false positive rate due to the fixed threshold combined with logical OR operation. The invention provides a promyelocytic leukemia identification device based on morphological image analysis, which can suppress normal background interference, focus on sparse pathological features, and quantify and coordinate the conflict between multi-scale evidence, thereby taking into account both the sensitivity and reliability of identification.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A device for identifying promyelocytic leukemia based on morphological image analysis, comprising: The multi-dimensional image acquisition and slice distribution module is used to acquire low-magnification images of blood smears under a low-magnification objective and high-magnification images under a high-magnification objective, and to divide the low-magnification images and the high-magnification images into image slices of a preset size, thereby obtaining a set of low-magnification slices and a set of high-magnification slices. An adaptive dual-channel morphological preprocessing module is used to convert the image slice from RGB color space to HSV color space. The saturation channel is thresholded using the maximum inter-class variance method to generate a binary saturation mask. The hue channel is filtered by hue range, Gaussian filtering, threshold binarization, and morphological opening to generate a binary hue mask. The saturation mask and the hue mask are combined by taking the Boolean intersection to obtain a white blood cell mask. The white blood cell mask is then multiplied pixel by pixel with the corresponding image slice to output a white blood cell target slice with the background set to zero. The cascaded residual multi-scale feature extraction module is used to extract features from low-magnification white blood cell target slices using a low-magnification encoder to obtain a low-magnification feature set, and to extract features from high-magnification white blood cell target slices using a high-magnification encoder to obtain a high-magnification feature set. The cross-scale gated attention multi-instance learning module is used to take the global average of each low-magnification feature in the low-magnification feature set to obtain a global query vector, use the global query vector as gating information to calculate the attention weight of each high-magnification feature in the high-magnification feature set, and perform a weighted summation of the high-magnification feature set based on the attention weight to obtain high-magnification aggregated features. The uncertainty resolution and fusion diagnosis module of evidence theory is used to obtain low-multiplied classification probabilities based on the global query vector and high-multiplied classification probabilities based on the high-multiplied aggregated features. It determines discount coefficients for the low-multiplied and high-multiplied classification probabilities based on their respective information entropies. Based on these discount coefficients, it reconstructs the low-multiplied and high-multiplied classification probabilities into basic probability assignments containing unknown states. It first calculates the conflict factor between the two basic probability assignments, and when the conflict factor is not greater than a conflict threshold, it uses the Dempster combination rule to fuse the two basic probability assignments. Based on the fusion result, it determines the confidence and likelihood for promyelocytic leukemia, and then outputs the identification result based on the confidence, likelihood, and conflict factor.

[0009] With the help of the above technical solutions, the cross-scale gating attention multi-instance learning module guides the weight allocation of high-magnification slices with macroscopic distribution information at low magnification scale, so that sparse pathological features can be focused without being diluted by normal background; the evidence theory uncertainty resolution and fusion diagnosis module measures the degree of uncertainty of the output of the two scales and coordinates the conflict, so that the decision no longer depends solely on a fixed threshold, thereby suppressing false positives while maintaining sensitivity.

[0010] As a preferred technical solution of the present invention, when the adaptive dual-channel morphological preprocessing module generates a tone mask, pixels in the tone channel that are within a preset hue range are marked as candidate foreground and the remaining pixels are marked as background to obtain an initial binary image. Gaussian filtering is applied to the initial binary image to smooth the edges and suppress high-frequency noise. Then, the result of Gaussian filtering is re-binarized with a preset judgment threshold to recover the binary mask. Subsequently, a morphological opening operation of erosion followed by dilation is performed with a structuring element of a preset size to remove isolated spots and cut off small adhesions to obtain a binary tone mask.

[0011] As a preferred technical solution of the present invention, when the adaptive dual-channel morphological preprocessing module generates a saturation mask, it iterates through the gray levels of the saturation channel to obtain a threshold gray level that maximizes the inter-class variance between the segmented foreground and background, and binarizes the saturation channel accordingly to obtain a binary saturation mask; when multiplying the white blood cell mask with the corresponding image slice pixel by pixel, the pixel values ​​outside the white blood cell mask are set to zero.

[0012] As a preferred embodiment of the present invention, both the low-magnification encoder and the high-magnification encoder use a residual network that integrates channel attention and spatial attention as the backbone, and use the features output by the global average pooling layer located before the fully connected layer in the residual network as the extraction result. Each feature in the low-magnification feature set and the high-magnification feature set is a feature vector of a preset dimension.

[0013] As a preferred embodiment of the present invention, the cross-scale gated attention multi-instance learning module calculates the global query vector according to the following formula:

[0014] in, The first in the low-magnification feature set A low-magnification feature, The number of low-magnification features in the set of low-magnification features.

[0015] As a preferred embodiment of the present invention, the cross-scale gated attention multi-instance learning module calculates the first element in the high-magnification feature set according to the following formula. Attention weights for high-magnification features:

[0016] in, The first in the high-magnification feature set A high-magnification feature, The number of high-magnification features in the high-magnification feature set. and These are the learnable linear mapping matrices acting on the high-magnification features and the global query vector, respectively. For learnable projection vectors, It is the hyperbolic tangent function. For the Sigmoid function, This indicates element-wise multiplication.

[0017] As a preferred embodiment of the present invention, the cross-scale gated attention multi-instance learning module performs a weighted summation of the high-magnification feature set according to the following formula to obtain the high-magnification aggregated features:

[0018] As a preferred embodiment of the present invention, the evidence theory uncertainty resolution and fusion diagnostic module uses two mutually exclusive propositions—promyelocytic leukemia and normal—to form an identification framework, and the low-power classification probability... middle and Let the probabilities of promyelocytic leukemia and normal leukemia be respectively. The information entropy is calculated using the following formula, and a discount factor is determined. Where a certain probability is zero, that term is considered... The value is zero:

[0019] Then, the low-power classification probabilities are reconstructed into basic probability assignments according to the following formula:

[0020] in, The unknown state corresponding to the identification framework is represented by the high-multiplier classification probability, which is reconstructed into the basic probability allocation in the same way.

[0021] As a preferred embodiment of the present invention, the evidence theory uncertainty resolution and fusion diagnosis module calculates the conflict factor according to the following formula:

[0022] When the conflict factor is greater than the conflict threshold, no fusion calculation is performed and a conflict warning is issued indicating that the case should be transferred to manual review; when the conflict factor is not greater than the conflict threshold, evidence fusion is performed according to the following formula, where the denominator is... Greater than zero:

[0023] in, , , All of these are propositions within the power set of the aforementioned identification framework. For any non-empty proposition that the desired fusion result is subject to, The proposition corresponding to the basic probability assignment obtained from the reconstruction of the low-power classification probabilities. The proposition corresponding to the basic probability assignment obtained from the reconstruction of the high-power classification probability. This represents the intersection of all sets. The proposition of Summation is performed; and the confidence and likelihood values ​​for promyelocytic leukemia are determined according to the following formula:

[0024] When the confidence level for promyelocytic leukemia is greater than the confidence threshold and the likelihood level is greater than the likelihood threshold, the output is a result identifying it as promyelocytic leukemia; when the confidence level for a normal proposition is greater than the confidence threshold, the output is a result identifying it as not promyelocytic leukemia; in other cases, the output is a result indicating suspicion and suggesting further testing.

[0025] This invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the following processing: acquiring a low-magnification image of a blood smear under a low-magnification objective lens and a high-magnification image under a high-magnification objective lens, and segmenting them into image slices of a preset size; converting each image slice from the RGB color space to the HSV color space; generating a binary saturation mask for the saturation channel using the maximum inter-class variance method; generating a binary tone mask for the tone channel using hue range filtering, Gaussian filtering, threshold binarization, and morphological opening operation; obtaining a white blood cell mask by taking the Boolean intersection of the two masks; and extracting white blood cell target slices with the background set to zero based on this mask; and performing separate processing on the low-magnification and high-magnification white blood cell target slices. Feature extraction yields a low-magnification feature set and a high-magnification feature set. A global average is taken from the low-magnification feature set to obtain a global query vector. Using this global query vector as gating information, attention weights are calculated for each high-magnification feature in the high-magnification feature set, and a weighted sum is obtained to obtain a high-magnification aggregated feature. Low-magnification classification probabilities and high-magnification classification probabilities are obtained based on the global query vector and the high-magnification aggregated feature, respectively. These probabilities are then reconstructed into a basic probability distribution containing unknown states based on the discount coefficient determined by their respective information entropies. A conflict factor is first calculated, and when the conflict factor is not greater than a conflict threshold, the Dempster combination rule is used for fusion. Based on the fusion result, the confidence level and likelihood for promyelocytic leukemia are determined, and the identification result is output.

[0026] Compared with the prior art, the present invention has the following beneficial effects: This invention obtains a global query vector by taking a global average of the low-magnification feature set, and uses this global query vector as gating information to participate in the calculation of attention weights for each high-magnification feature. This results in high-magnification sections containing pathological features such as abnormal particles and irregular karyotypes being given higher weights, while the weights of normal cells or background sections are correspondingly reduced. Based on this, the high-magnification features are weighted and summed to obtain aggregated features, thereby avoiding the problem of sparse pathological features being diluted by normal background due to indiscriminately taking an arithmetic average of the probabilities of each section. Since the allocation of high-magnification section weights is guided by the macroscopic distribution information at the low-magnification scale, a contextual association is established between the low-magnification and high-magnification scales. This allows even if a randomly selected section contains only a small number of abnormal cells, its corresponding features can still be focused and amplified, which is beneficial to improving the sensitivity of recognition.

[0027] This invention measures the uncertainty of each output by calculating the information entropy of low-magnification and high-magnification classification probabilities separately, and determines the discount coefficient accordingly. The classification probabilities are then reconstructed into basic probability assignments containing unknown states. Dempster's combination rule is then used for fusion and conflict factor calculation. When the probability output by one scale is too high due to staining defects or preparation artifacts, which contradicts the strong negative information given by another scale, the conflict factor increases accordingly. The fusion process coordinates the conflict between the two sources of evidence, so that the decision no longer depends solely on whether a single scale exceeds a fixed threshold. This reduces false positives caused by local noise or artifacts while maintaining the sensitivity of promyelocytic leukemia identification, and alleviates the burden of subsequent manual review.

[0028] Because the cross-scale gated attention multi-instance learning module provides features and classification probabilities with high signal-to-noise ratio for subsequent fusion, and the evidence theory uncertainty resolution and fusion diagnosis module coordinates the unavoidable statistical noise and scale contradictions in the previous stage, the two work together to enable the device to balance the sensitivity and reliability of recognition when facing blood smear images with interference factors such as differences in staining batches, uneven coating, and overlapping of red blood cells and platelets, and to output conflict warnings when there are serious conflicts of evidence so as to transfer to manual review. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the overall module structure of the promyelocytic leukemia identification device in an embodiment of the present invention; Figure 2 This is a schematic diagram of the processing flow of the adaptive dual-channel morphological preprocessing module in an embodiment of the present invention; Figure 3 This is a schematic diagram of the data flow of the cross-scale gated attention multi-instance learning module in an embodiment of the present invention; Figure 4 This is a schematic diagram of the decision logic flow of the evidence theory uncertainty resolution and fusion diagnosis module in an embodiment of the present invention.

[0030] Figure labeling: 101, Multi-dimensional image acquisition and slice distribution module; 102, Adaptive dual-channel morphological preprocessing module; 103, Cascaded residual multi-scale feature extraction module; 104, Cross-scale gated attention multi-instance learning module; 105, Evidence theory uncertainty resolution and fusion diagnosis module. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0032] The promyelocytic leukemia identification device based on morphological image analysis disclosed in this invention comprises components deployed in a software pipeline on a server with graphics processor acceleration capabilities or a computer-readable storage medium storing corresponding computer programs, working collaboratively according to the order of data flow. For example... Figure 1 As shown, the device sequentially includes a multi-dimensional image acquisition and slice distribution module 101, an adaptive dual-channel morphological preprocessing module 102, a cascaded residual multi-scale feature extraction module 103, a cross-scale gated attention multi-instance learning module 104, and an evidence theory uncertainty resolution and fusion diagnosis module 105. Intermediate data is passed between adjacent modules in tensor form, and the output of the previous module constitutes the input of the next module.

[0033] The multi-dimensional image acquisition and slice distribution module 101 is used to receive and process image data from the digital microscopy scanning device. When the blood smear to be examined is placed on the stage of the automated microscopy scanning platform, the microscopy system sequentially uses low-power and high-power objectives to digitally scan the entire smear or a single-layer area with uniformly distributed cells, obtaining two sets of full-field digital images: low-power and high-power. Considering that the size of full-field digital images is usually extremely large, the multi-dimensional image acquisition and slice distribution module 101 crops them into standard-sized image blocks using an overlapping sliding window method according to a preset step size. In this embodiment, the standard size is taken as... The overlap ratio of the sliding windows is set to 10%, meaning the step size of adjacent windows is 720 pixels, to avoid damaging cells located at the cutting boundaries. Therefore, the multi-dimensional image acquisition and slice distribution module 101 segments the low-magnification image into a set of low-magnification slices. The high-magnification image is segmented into a set of high-magnification slices. ,in and These represent the number of sections at low and high magnification, respectively. Because the physical area per unit area is smaller and the sections are denser at high magnification, there are typically [number missing]. Taking a specific application example, for a single-layer region of a peripheral blood smear, the multi-dimensional image acquisition and slice distribution module 101 crops the image from a low-magnification image. A slice, cropped from a high-magnification image. Each slice corresponds to a size of [number]. The image tensor. The multi-dimensional image acquisition and slice distribution module 101 then distributes the above slice set as a tensor to the cache queue of the downstream module.

[0034] Blood smears contain a large number of red blood cells and platelets, which can easily interfere with the extraction of white blood cell morphology. Therefore, the adaptive dual-channel morphological preprocessing module 102 performs white blood cell target separation on each image slice. The processing flow is described in [reference needed]. Figure 2 The adaptive dual-channel morphological preprocessing module 102 first converts each incoming RGB three-channel image slice from the RGB color space to the HSV color space after photometric calibration. For slices normalized to... The red, green, and blue components of the interval , , Its lightness and saturation components are calculated according to the following formulas:

[0035]

[0036] Tone Components The conversion is then performed according to the standard segmentation relationship based on which channel yields the maximum value among the three components. This conversion decouples brightness and chroma to some extent, resulting in a higher contrast in the saturation and hue channels for leukocyte nuclei and cytoplasm, which appear purplish-blue after Wright-Gymsa staining, compared to erythrocytes and plasma. Subsequently, the adaptive dual-channel morphological preprocessing module 102 extracts the saturation channel data and uses the one-dimensional Otsu's method to search across all possible gray levels. The gray level is denoted as... When the boundary is set, the probabilities of being classified into foreground and background are respectively... and The mean gray values ​​of the two classes are respectively and Then the inter-class variance and the optimal threshold Determined according to the following formula:

[0037] In the aforementioned application example, for quantizing the saturation channel to... In this case, the optimal threshold obtained by the adaptive dual-channel morphological preprocessing module 102 is approximately Based on this, the highly saturated white blood cell nuclei and granules are classified as the foreground, resulting in a binary saturation mask where the pixel value is only 0 or 1. Meanwhile, the adaptive dual-channel morphological preprocessing module 102 extracts data from the hue channel, ensuring that the hue values ​​fall within a preset hue range (in this embodiment, the normalized hue range is approximately...). Pixels within the region corresponding to the purplish-blue phase of the stained cell nucleus are labeled as candidate foreground, and the remaining pixels are labeled as background to obtain an initial binary image. To suppress possible residual platelet artifacts and high-frequency noise caused by red blood cell edge diffraction within this region, the adaptive dual-channel morphological preprocessing module 102 first applies a two-dimensional Gaussian smoothing filter to the initial binary image. In this embodiment, the Gaussian kernel size used is [missing information]. The standard deviation is approximately 1.0. Considering that edge pixels will degenerate into continuous gray values ​​between 0 and 1 after applying Gaussian filtering to the binary image, the adaptive dual-channel morphological preprocessing module 102 then re-binarizes the Gaussian filtering result with a preset judgment threshold (0.5 in this embodiment), restoring it to a binary image where pixel values ​​are only 0 or 1. Then, it concatenates the image again, with a structuring element size of... The binary mathematical morphological opening operation of pixels, namely erosion followed by dilation, is used to cut off small, adhered areas and remove isolated small speckle noise, ultimately resulting in a binary tone mask. Due to saturation mask The tone mask is obtained by thresholding using the maximum inter-class variance method. After the above re-binarization process, both are binary masks with pixel values ​​of only 0 or 1. The adaptive dual-channel morphological preprocessing module 102 then performs matrix-level Boolean intersection operations accordingly. This operation can be reliably implemented in the program using a matrix bitwise AND operation, resulting in a white blood cell mask that can relatively completely encompass the outline of white blood cells. Finally, the adaptive dual-channel morphological preprocessing module 102 masks the white blood cells. According to the corresponding original RGB image slices Pixel-by-pixel multiplication is performed, and all pixel values ​​in the background region outside the white blood cell mask are set to zero. The output is a slice of the white blood cell target with only the white blood cell target and a black background, which is then fed into the next level module. Compared with the method of segmentation using a single threshold on a single channel, the adaptive dual-channel morphological preprocessing module 102 provides constraints for saturation and hue channels respectively and takes their intersection. This can overcome the shortcomings of single channels being greatly affected by staining color differences to a certain extent, resulting in a purer white blood cell target, thereby reducing the possibility of the subsequent network learning invalid background features.

[0038] The cascaded residual multi-scale feature extraction module 103 instantiates the low-magnification encoder in parallel on the graphics processor computing node. With high magnification encoder Two deep convolutional neural network branches, both structurally employing residual networks that integrate channel attention and spatial attention mechanisms as their backbones; in this embodiment, ResNet50 can be used. To obtain highly condensed deep feature representations that are not yet compressed into specific categories, the cascaded residual multi-scale feature extraction module 103, during network forward propagation, does not pass through the fully connected layer and classification layer at the end of the residual network, but instead extracts the features output by the global average pooling layer located before the fully connected layer. Thus, the cascaded residual multi-scale feature extraction module 103 extracts a low-magnification feature set from the low-magnification white blood cell target slice. High-magnification feature sets were extracted from high-magnification white blood cell target slices. Each feature and All A 1-dimensional real vector This is the feature dimension of the residual network output layer, typically taken as 1024 or 2048. In this embodiment, it is taken as... Following the aforementioned application example, the cascaded residual multi-scale feature extraction module 103 outputs a low-magnification feature set consisting of four 2048-dimensional vectors and a high-magnification feature set consisting of twenty 2048-dimensional vectors, and sends them to the next-level module in batches.

[0039] The cross-scale gated attention multi-instance learning module 104 is used to dynamically focus discriminative local features from high-magnification slices containing a large background of normal cells, with the data flow as follows: Figure 3 As shown in the multi-instance learning framework, the entire smear or a specific pathological region within it is considered a package, and the high-power sections segmented from it are considered instances within that package. Traditional methods aggregate the instance features using operators such as max pooling or average pooling. The former is prone to losing information, while the latter tends to dilute sparse key features. It is worth noting that if the arithmetic mean of the predicted probabilities for each high-power section is used, in the aforementioned application example, assuming that only 2 out of 20 high-power sections contain typical bundled cells and are given a high disease probability of approximately 0.95 by the network, while the remaining 18 normal sections are given a low probability of approximately 0.05, then the arithmetic mean is only... This value is below the commonly used judgment threshold of 0.3, making it highly susceptible to misclassification as a non-case and resulting in missed diagnosis; this is precisely the consequence of sparse pathological features being diluted by the normal background during the arithmetic averaging process. Therefore, the cross-scale gated attention multi-instance learning module 104 first processes the low-magnification feature set... A global average is performed on all low-magnification features in the sample to obtain a global query vector that characterizes the overall cell distribution density and proliferation trend of the blood smear under low-magnification field of view. :

[0040] Subsequently, the cross-scale gated attention multi-instance learning module 104 uses the global query vector As gating information, for high-magnification feature sets For each high-magnification feature in the matrix, an attention weight is calculated. High magnification features Its attention weight Determined according to the following formula:

[0041] In the formula, and The high-magnification features are mapped to the global query vector to a dimension of 1. The learnable linear mapping matrix of the hidden layer, To project the hidden layer features into learnable weight vectors for the attention scalar, in this embodiment, the hidden layer dimension is taken as... , The hyperbolic tangent function is used to constrain the representation range of features to... Inside, The sigmoid function constrains the values ​​of the global query vector after mapping to... This constitutes the gating information. This indicates element-wise multiplication. As can be seen from this formula, the importance assessment of high-magnification features is not solely determined by the features themselves, but is influenced by the global query vector from the low-magnification scale. Modulation: When the microstructure reflected by a high-magnification feature matches the anomalous macroscopic distribution reflected by a low-magnification scale, its corresponding attention weight... Amplified values ​​are magnified, while suppressed values ​​are reduced and approach zero. In the aforementioned application example, after calculation by the cross-scale gated attention multi-instance learning module 104, the attention weights corresponding to the two high-magnification slices containing typical bundle cells are approximately 0.45 and 0.37, respectively, totaling approximately 0.82. The sum of the attention weights corresponding to the remaining 18 normal slices is only approximately 0.18, averaging approximately 0.01 per slice. All weights, after Softmax normalization, satisfy... After obtaining the attention weights for each high-magnification slice, the cross-scale gated attention multi-instance learning module performs a multi-instance learning of 104 high-magnification feature sets. Performing linear weighted summation yields high-magnification aggregated features that represent the pathological information of the smear at the microscopic level. :

[0042] Thus, the originally numerous and redundant high-magnification features are aggregated into a single high-magnification aggregated feature. Features contributed by a few sections containing typical lesions are highlighted due to being assigned a cumulative weight of approximately 0.82, while the contributions of normal cells and background sections are correspondingly suppressed; compared to the aforementioned case where the arithmetic mean is diluted to 0.14, this high-magnification aggregated feature... This allows key information from sparse positive sections to be preserved.

[0043] The Evidence Theory Uncertainty Resolution and Fusion Diagnosis Module 105 uses global query vectors from low-magnification scales. High-magnification aggregation features from high-magnification scale As two relatively independent sources of evidence, the judgment logic process is as follows: Figure 4 As shown. The uncertainty resolution and fusion diagnostic module 105 constructs a low-magnification classifier in parallel. With high-magnification classifier , global query vector Input low-magnification classifier Obtain low-power classification probability High-magnification aggregation features Input high-magnification classifier Obtain high classification probability ,in , These represent the predicted probabilities of promyelocytic leukemia at the corresponding scales. , These represent the predicted probabilities of normality at the corresponding scales. The uncertainty resolution and fusion diagnostic module 105 of the evidence theory defines the identification framework for medical identification as a set of mutually exclusive propositions. Its power set is , of which elements That is, the corresponding identification framework It is used to represent an ambiguous or unknown state in the process of identification.

[0044] Considering that the probability output by the classifier after normalization satisfies However, it cannot express the model's lack of confidence in its own judgment. The uncertainty resolution and fusion diagnosis module 105 of evidence theory introduces Shannon information entropy as a measure of uncertainty. Taking low-magnification scale as an example, let... and Let be the probabilities of promyelocytic leukemia and normal leukemia, respectively. Then, the information entropy of their classification probability distributions is... Calculate according to the following formula:

[0045] Since the identification frame contains only two mutually exclusive propositions, the base-2 information entropy falls within the range of... Within the given interval, when both class probabilities are close to 0.5, the information entropy approaches its maximum value of 1, indicating that the model is in a highly conflicted state; when one class probability is close to 1 and the other is close to 0, the information entropy approaches 0, indicating that the model has made a relatively definite judgment. It should be noted that, according to the conventions of information theory, when one class probability is zero, it is agreed that... Therefore, the aforementioned information entropy is bounded and well-defined across the entire range of the probability distribution. In engineering implementation, to avoid the probability output by the classifier via Softmax taking an exact zero during floating-point underflow or rounding, ... If the divergence is negative infinity, a very small positive constant much smaller than the probability resolution can be superimposed on each probability before taking the logarithm. (For example, take) ), that is, with The calculation is performed without affecting the effective value of the entropy. Accordingly, the evidence theory uncertainty resolution and fusion diagnosis module 105 determines the discount coefficient using normalized information entropy. The discount factor decreases as the model becomes more entangled with the recognition results and the information entropy increases. The smaller the value, the better. Furthermore, the uncertainty resolution and fusion diagnostic module 105 of the evidence theory reconstructs the classification probabilities of low-magnification evidence sources into basic probability assignments that include unknown states:

[0046]

[0047]

[0048] The portion of probability quality deducted by the discount factor All were transferred to represent an unknown state. This explicitly incorporates the uncertainty of the model into the expression of evidence. Basic probability allocation of high-magnification scale evidence sources. , and The information entropy of the high-power classification probability distribution is calculated in parallel in exactly the same way as above, where it is denoted as... The discount factor is denoted as .

[0049] After obtaining the two independent basic probability assignments and Subsequently, the uncertainty resolution and fusion diagnosis module 105 of the evidence theory first calculates the conflict factor used to measure the degree of contradiction between the two pieces of evidence. It is equal to the sum of the mass products of all proposition pairs whose intersection is an empty set, specifically as follows within this identification framework:

[0050] In the above formula and subsequent fusion formulas, and These are the propositions corresponding to the basic probability assignments at the low and high magnification scales, respectively, both of which are taken from the power set of the identification framework. Denotes a pair of propositions whose intersection is an empty set. Summation. When one piece of evidence, due to severe slide contamination, is judged to be normal with a higher probability, while another piece of evidence is judged to be promyelocytic leukemia with a higher probability, the intersection of the two propositions is an empty set, and the conflict factor is... The conflict factor increases accordingly. The uncertainty resolution and fusion diagnosis module 105 of the evidence theory first determines the timing of subsequent algorithms based on the comparison results of the conflict factor and the conflict threshold: when the conflict factor... When the conflict factor exceeds the conflict threshold, it indicates that the two pieces of evidence are severely contradictory. At this point, subsequent fusion calculations are no longer performed. Instead, a conflict warning is directly output and the process is transferred to manual review. This avoids, from an algorithmic timing perspective, giving contradictory conclusions and conflict factors with extremely high confidence on both scales. When the denominator of the fusion formula approaches 1 The division-by-zero anomaly caused by approaching zero; only when the conflict factor... When the conflict threshold is not greater than 1, the Evidence Theory Uncertainty Resolution and Fusion Diagnosis Module 105 uses the Dempster combination rule to fuse the two pieces of evidence. Since the conflict threshold is less than 1, the denominator is... Strictly greater than zero, fusion computation is well-defined. For any non-empty proposition in the identification framework. Basic probability allocation after fusion Calculate according to the following formula:

[0051] In the fused denominator This serves a normalization function, redistributing the probabilistic mass lost due to conflicting evidence to non-empty propositions, thus achieving the dual effect of smoothing noise and fusing complementary evidence. Building upon this, the uncertainty resolution and fusion diagnosis module 105 of evidence theory further derives the confidence and likelihood of the proposition regarding promyelocytic leukemia. Confidence is the lower bound of confirmation, equal to the mass allocated to the proposition itself after fusion; likelihood is the upper bound of non-confrontation, equal to the sum of the mass allocated to the proposition itself after fusion and the mass allocated to the unknown state.

[0052]

[0053] Difference between trust level and likelihood It quantitatively reflects the ambiguous and unknown aspects that cannot be eliminated in this identification process.

[0054] To more clearly illustrate the operational process of the uncertainty resolution and fusion diagnostic module 105 in evidence theory and its difference from the fixed threshold combined logical OR operation, several specific numerical examples are provided below. In the case where both scales consistently point to a positive result, continuing from the aforementioned application example, assuming the low-power classification probability is... The probability of high-power classification is Then the information entropy at a low scale Discount coefficient The reconstructed basic probability allocation is as follows: , , Information entropy at high scale Discount coefficient The reconstructed basic probability allocation is as follows: , , Calculate the conflict factor based on this. Since its value is not greater than the conflict threshold, the Evidence Theory Uncertainty Resolution and Fusion Diagnosis Module 105 continues to perform fusion calculations, and the result obtained after fusion using the Dempster combination rule is... , , Thus, trust Likelihood Since both the confidence level and likelihood exceed the corresponding thresholds, the uncertainty resolution and fusion diagnosis module 105 outputs a diagnosis of promyelocytic leukemia. In the case where both scales consistently point to a negative result, assuming the low-power classification probability is... The probability of high-power classification is The conflict factor can be obtained by following the same process. Trust level regarding normal propositions Based on this, the uncertainty resolution and fusion diagnosis module 105 outputs the identification result of non-promyelocytic leukemia.

[0055] The advantages of this invention are particularly evident when evidence from two scales contradicts each other. It assumes that the high-magnification scale incorrectly gives an inflated probability of disease due to staining defects in the sections. However, low-magnification scales provide strong evidence of dissent based on a macroscopic perspective. If a fixed threshold combined with a logical OR operation is used, the system will be forced to classify a high-magnification probability of 0.90 as positive, resulting in a false positive, since this probability exceeds the threshold of 0.3. With this invention, low-magnification reconstruction yields... , , High-scale reconstruction yielded , , Conflict factors The conflict threshold was not exceeded, so fusion continued. The confidence level regarding the proposition concerning promyelocytic leukemia after fusion was assessed. The confidence threshold was not reached, therefore no positive conclusion was output. Instead, it fell into a borderline state of weak uncertainty and was treated as a suspected case, thus effectively suppressing false positives caused by local noise. Furthermore, when both scales give opposite conclusions with extremely high confidence, for example, when the low-power classification probability is... The probability of high-power classification is At this time, the conflict factor The conflict threshold has been exceeded. Based on this, the Evidence Theory Uncertainty Resolution and Fusion Diagnosis Module 105 directly outputs a conflict warning and transfers it to the manual review process before performing fusion calculation. This avoids forcibly giving potentially erroneous conclusions when the evidence is severely inverted, and also avoids the possible operational anomalies caused by performing fusion division when the conflict factor is close to 1.

[0056] The Evidence Theory Uncertainty Resolution and Fusion Diagnosis Module 105 performs multi-branch logical decisions to output identification results. In this embodiment, based on empirical thresholds determined after verification from a large number of historical cases, the conflict threshold is set to 0.70, the trust threshold to 0.60, and the likelihood threshold to 0.85. The Evidence Theory Uncertainty Resolution and Fusion Diagnosis Module 105 calculates the conflict factor... First, conflict determination is performed: when the conflict factor... When the conflict threshold is exceeded, it indicates a serious logical inversion between the evidence at the two scales, possibly due to partial slide contamination or mismatched labeling. The Evidence Theory Uncertainty Resolution and Fusion Diagnosis Module 105 will no longer perform fusion calculations, directly outputting a conflict warning, refusing to force a conclusion, and transferring the process to manual microscopy review; only when the conflict factor... Evidence fusion is only performed when the conflict threshold is not exceeded, and subsequent judgments are made based on the fusion results: when the confidence level of the proposition regarding promyelocytic leukemia is... Greater than the trust threshold and likelihood When the likelihood threshold is greater than 105, the uncertainty resolution and fusion diagnosis module 105 outputs a diagnosis of promyelocytic leukemia; when the confidence level of the normal proposition is greater than 105%, the probability of the normal proposition is greater than 105%. When the threshold is greater than the confidence threshold, the Evidence Theory Uncertainty Resolution and Fusion Diagnostic Module 105 outputs a result indicating non-promyelocytic leukemia. In other cases, the Evidence Theory Uncertainty Resolution and Fusion Diagnostic Module 105 outputs a result indicating a weakly uncertain borderline state, classifying the case as a suspected case, and recommending further confirmation using flow cytometry and relevant gene testing indicators. The aforementioned thresholds can be adjusted according to the different emphases on sensitivity and specificity in the screening scenario. When the emphasis is on avoiding missed detections, the confidence threshold can be appropriately lowered; when the emphasis is on reducing false positives, the confidence threshold and likelihood threshold can be appropriately increased.

[0057] To facilitate understanding of the complete collaborative process of the various modules in this device, its overall operation flow is described below. After the peripheral blood or bone marrow smear submitted for clinical testing is prepared, the operator places it into the automated digital pathology scanner. Under the command of the main control computer, the scanner automatically focuses and scans in a grid-like manner. First, a panoramic image is acquired through a low-power objective lens, and then the image is switched to a high-power objective lens to continuously acquire a high-resolution field of view in areas with suitable cell distribution. The obtained image is written to a high-speed solid-state storage medium mounted on the server via a high-bandwidth link. The multi-dimensional image acquisition and slice distribution module 101 reads and trims the slice tensors to standard sizes in a multi-threaded manner. Each slice tensor enters the adaptive dual-channel morphological preprocessing module 102, where the conversion from RGB color space to HSV color space, as well as the mask calculation, re-binarization, and intersection calculation of the saturation and hue channels, are performed in parallel on the graphics processor, outputting a white blood cell target slice with a zero background. The purified slice tensors are packaged and sent to the low-power encoder in the cascaded residual multi-scale feature extraction module 103. With high magnification encoder As the network depth increases, the spatial dimension is gradually downsampled while the channel dimension gradually expands. The channel and spatial attention mechanisms interspersed throughout the network continuously enhance the response to structures sensitive to particles, kernel folds, etc. Finally, each slice is encoded into a feature vector of fixed dimension in the global average pooling layer. The above feature flow enters the cross-scale gated attention multi-instance learning module 104, where a global query vector is synthesized from low-magnification features. This information is used as gating information to apply cross-scale attention weighting to each high-magnification feature, amplifying the weights of a few high-magnification sections containing typical lesions while suppressing the weights of the remaining sections, thus creating a high-magnification feature set. Therefore, it is aggregated into a high-multiplication aggregation feature. Finally, the global query vector. With high-magnification aggregation characteristics Simultaneously, the module 105 for resolving uncertainty in evidence theory and fusion diagnosis obtains low-magnification and high-magnification classification probabilities, determines discount coefficients through information entropy, and reconstructs basic probability allocations containing unknown states. First, the conflict factor is calculated and conflict judgment is performed. When the conflict factor is not greater than the conflict threshold, the Dempster combination rule is used to fuse and obtain the confidence and likelihood. Then, multi-branch logic decision is executed to give the recognition result or issue an early warning to transfer to manual review when there is serious evidence conflict. This completes the complete closed loop from image input, target purification, feature representation, cross-scale attention aggregation to evidence fusion decision.

[0058] Without departing from the core concept of this invention, those skilled in the art can make various equivalent modifications to the above embodiments. For example, the standard size of the image slices in the multi-dimensional image acquisition and slice distribution module 101 is not limited to... The overlap ratio of the sliding window and the number of low-magnification and high-magnification slices can be appropriately increased or decreased according to cell density and computing resources; the size of the structural element used in the morphological opening operation in the adaptive dual-channel morphological preprocessing module 102 can be adjusted. The pixel area is adjusted appropriately based on noise and target scale. The kernel size and variance of the Gaussian filter, the decision threshold used for re-binarization, and the upper and lower limits of the hue range can also be set accordingly. The coloring method used is not limited to Wright-Gymsa staining. The backbone network in the cascaded residual multi-scale feature extraction module 103 is not limited to ResNet50. Any convolutional neural network that can extract multi-scale deep features and fuse channel and spatial attention can be used. The truncated feature dimension is... It is not limited to 1024 or 2048; the hidden layer dimensions in the cross-scale gated attention multi-instance learning module are 104. The activation function in the gated attention can be set according to the feature dimension and sample size, while maintaining the gating characteristics. Other equivalent forms can be adopted. The identification framework in the evidence theory uncertainty resolution and fusion diagnosis module 105 can be expanded accordingly when more cell categories need to be distinguished. In addition to using information entropy with base 2, other normalized uncertainty measures can also be used to measure uncertainty. The conflict threshold, trust threshold, and likelihood threshold used in multi-branch decision can be calibrated according to the specific application scenario and different requirements for sensitivity and specificity. All of these are natural extensions and refinements of the technical solution of this invention and still fall within the protection scope of this invention.

[0059] Furthermore, the promyelocytic leukemia identification device based on morphological image analysis disclosed in this invention, whose processing functions include the multi-dimensional image acquisition and slice distribution module 101, the adaptive dual-channel morphological preprocessing module 102, the cascaded residual multi-scale feature extraction module 103, the cross-scale gated attention multi-instance learning module 104, and the evidence theory uncertainty resolution and fusion diagnosis module 105, can be implemented by a computer program stored on a computer-readable storage medium. When the computer program is executed by a processor, the above-mentioned processing is completed. The computer-readable storage medium includes, but is not limited to, read-only memory, random access memory, disk, optical disk, and solid-state memory, etc., which are media capable of storing program code.

[0060] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of protection of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A device for identifying promyelocytic leukemia based on morphological image analysis, characterized in that, include: The multi-dimensional image acquisition and slice distribution module is used to acquire low-magnification images of blood smears under a low-magnification objective and high-magnification images under a high-magnification objective, and to divide the low-magnification images and the high-magnification images into image slices of a preset size, thereby obtaining a set of low-magnification slices and a set of high-magnification slices. An adaptive dual-channel morphological preprocessing module is used to convert the image slice from RGB color space to HSV color space. The saturation channel is thresholded using the maximum inter-class variance method to generate a binary saturation mask. The hue channel is filtered by hue range, Gaussian filtering, threshold binarization, and morphological opening to generate a binary hue mask. The saturation mask and the hue mask are combined by taking the Boolean intersection to obtain a white blood cell mask. The white blood cell mask is then multiplied pixel by pixel with the corresponding image slice to output a white blood cell target slice with the background set to zero. The cascaded residual multi-scale feature extraction module is used to extract features from low-magnification white blood cell target slices using a low-magnification encoder to obtain a low-magnification feature set, and to extract features from high-magnification white blood cell target slices using a high-magnification encoder to obtain a high-magnification feature set. The cross-scale gated attention multi-instance learning module is used to take the global average of each low-magnification feature in the low-magnification feature set to obtain a global query vector, use the global query vector as gating information to calculate the attention weight of each high-magnification feature in the high-magnification feature set, and perform a weighted summation of the high-magnification feature set based on the attention weight to obtain high-magnification aggregated features. The uncertainty resolution and fusion diagnosis module of evidence theory is used to obtain low-multiplied classification probabilities based on the global query vector and high-multiplied classification probabilities based on the high-multiplied aggregated features. It determines discount coefficients for the low-multiplied and high-multiplied classification probabilities based on their respective information entropies. Based on these discount coefficients, it reconstructs the low-multiplied and high-multiplied classification probabilities into basic probability assignments containing unknown states. It first calculates the conflict factor between the two basic probability assignments, and when the conflict factor is not greater than a conflict threshold, it uses the Dempster combination rule to fuse the two basic probability assignments. Based on the fusion result, it determines the confidence and likelihood for promyelocytic leukemia, and then outputs the identification result based on the confidence, likelihood, and conflict factor.

2. The promyelocytic leukemia identification device based on morphological image analysis according to claim 1, characterized in that, When the adaptive dual-channel morphological preprocessing module generates a tone mask, it marks pixels in the tone channel that are within a preset hue range as candidate foreground and the remaining pixels as background to obtain an initial binary image. Gaussian filtering is applied to the initial binary image to smooth the edges and suppress high-frequency noise. Then, the result of Gaussian filtering is re-binarized with a preset judgment threshold to recover the binary mask. Subsequently, a morphological opening operation of erosion followed by dilation is performed with a structuring element of a preset size to remove isolated spots and cut off small adhesions, thus obtaining a binary tone mask.

3. The promyelocytic leukemia identification device based on morphological image analysis according to claim 1, characterized in that, When the adaptive dual-channel morphological preprocessing module generates a saturation mask, it iterates through the gray levels of the saturation channels to obtain a threshold gray level that maximizes the inter-class variance between the segmented foreground and background, and then binarizes the saturation channels accordingly to obtain a binary saturation mask. When multiplying the white blood cell mask pixel by pixel with the corresponding image slice, the pixel values ​​outside the white blood cell mask are set to zero.

4. The promyelocytic leukemia identification device based on morphological image analysis according to claim 1, characterized in that, Both the low-magnification encoder and the high-magnification encoder use a residual network that integrates channel attention and spatial attention as the backbone, and use the features output by the global average pooling layer located before the fully connected layer in the residual network as the extraction result. Each feature in the low-magnification feature set and the high-magnification feature set is a feature vector of a preset dimension.

5. The promyelocytic leukemia identification device based on morphological image analysis according to claim 1, characterized in that, The cross-scale gated attention multi-instance learning module calculates the global query vector according to the following formula: ; in, The first in the low-magnification feature set A low-magnification feature, The number of low-magnification features in the set of low-magnification features.

6. The promyelocytic leukemia identification device based on morphological image analysis according to claim 5, characterized in that, The cross-scale gated attention multi-instance learning module calculates the first [number] feature in the high-magnification feature set according to the following formula. Attention weights for high-magnification features: ; in, The first in the high-magnification feature set A high-magnification feature, The number of high-magnification features in the high-magnification feature set. and These are the learnable linear mapping matrices acting on the high-magnification features and the global query vector, respectively. For learnable projection vectors, It is the hyperbolic tangent function. For the Sigmoid function, This indicates element-wise multiplication.

7. The promyelocytic leukemia identification device based on morphological image analysis according to claim 6, characterized in that, The cross-scale gated attention multi-instance learning module performs a weighted summation of the high-magnification feature set according to the following formula to obtain the high-magnification aggregated features: 。 8. The promyelocytic leukemia identification device based on morphological image analysis according to claim 1, characterized in that, The uncertainty resolution and fusion diagnostic module of the evidence theory uses the two mutually exclusive propositions of promyelocytic leukemia and normality to form an identification framework, and the low-power classification probability... middle and The probabilities of promyelocytic leukemia and normal status are respectively given. The information entropy of the low-power classification probabilities is calculated using the following formula, and a discount factor is determined. Where a certain class probability is zero, this term is defined as... The value is zero: Then, the low-power classification probabilities are reconstructed into basic probability assignments according to the following formula: ; in, The unknown state corresponding to the identification framework is represented by the high-multiplier classification probability, which is reconstructed into the basic probability allocation in the same way.

9. The promyelocytic leukemia identification device based on morphological image analysis according to claim 8, characterized in that, The uncertainty resolution and fusion diagnosis module of the evidence theory calculates the conflict factor according to the following formula: ; When the conflict factor is greater than the conflict threshold, the following fusion calculation is not performed, and a conflict warning for manual review is directly output; when the conflict factor is not greater than the conflict threshold, the two basic probability allocations are fused according to the following formula, where the denominator is... Greater than zero: ; in, , , All of these are propositions within the power set of the aforementioned identification framework. For any non-empty proposition that the desired fusion result pertains to, The proposition corresponding to the basic probability assignment obtained from the reconstruction of the low-power classification probabilities. The proposition corresponding to the basic probability assignment obtained from the reconstruction of the high-power classification probability. This represents the intersection of all sets. The proposition of Summation is performed; and the confidence and likelihood values ​​for promyelocytic leukemia are determined according to the following formula: ; When the confidence level for promyelocytic leukemia is greater than the confidence threshold and the likelihood level is greater than the likelihood threshold, the output is a result identifying it as promyelocytic leukemia; when the confidence level for a normal proposition is greater than the confidence threshold, the output is a result identifying it as not promyelocytic leukemia. In other cases, the system outputs a suspected result and suggests further testing.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it performs the following processing: acquiring low-magnification images of blood smears under a low-magnification objective lens and high-magnification images under a high-magnification objective lens, and segmenting them into image slices of preset sizes; converting each image slice from RGB color space to HSV color space; generating a binary saturation mask for the saturation channel using the maximum inter-class variance method; generating a binary tone mask for the tone channel using hue range filtering, Gaussian filtering, threshold binarization, and morphological opening operations; obtaining a white blood cell mask by taking the Boolean intersection of the two masks; and extracting white blood cell target slices with zero background based on this mask; and performing feature extraction on the low-magnification and high-magnification white blood cell target slices respectively to obtain a low-magnification feature set and a high-magnification feature set. The low-magnification feature set is used as a global feature vector. The global query vector is obtained by taking the global average of the low-magnification feature set. The global query vector is used as the gating information to calculate the attention weight of each high-magnification feature in the high-magnification feature set and the weighted sum is used to obtain the high-magnification aggregated feature. The low-magnification classification probability and the high-magnification classification probability are obtained according to the global query vector and the high-magnification aggregated feature, respectively. They are reconstructed into a basic probability allocation containing unknown states according to the discount coefficient determined by their respective information entropy. The conflict factor is calculated first, and when the conflict factor is not greater than the conflict threshold, the Dempster combination rule is used for fusion. The trust and likelihood for promyelocytic leukemia are determined according to the fusion result, and the recognition result is output.

Images