Explainable recognition system based on multiscale probability mapping morphological feature quantification

Abstract

The case relates to an interpretable recognition system based on multi-scale probability mapping morphological feature quantization, which is used to solve the lack of objective quantitative information basis for nodule diagnosis in the prior art, and the ambiguity information and diagnosis error in consultation communication due to the semantic ambiguity of qualitative concept factors. The system learns multi-scale feature distribution, extracts lesion contour, estimates uncertain error, quantizes morphological features, and recognizes the interpretability of the lesion, effectively realizes the automation and standardization of quantitative analysis of lesions in ultrasound images, improves the accuracy and interpretability of lesion recognition, and has important application value for quantitative analysis of 2D ultrasound images and assisting clinicians in formulating accurate diagnosis and treatment schemes. Further, the system designs a mask convolution guided hybrid auto-encoding self-supervised learning model in model training and learning, reduces the workload of manual annotation, and improves the accurate expression of ultrasonic lesion morphological features.

Description

Technical Field

[0001] This case involves technical fields such as computer vision and artificial intelligence, and medical image processing, especially interpretable recognition systems based on multi-scale probability mapping and morphological feature quantification. Background Technology

[0002] In medical image analysis, doctors often perform simple localization and scale measurements of regions of interest. Due to the high complexity of morphological features and the difficulty of quantitative analysis, doctors often use simplified conceptual descriptors to qualitatively describe the morphological characteristic factors of lesions. Although this approach improves the interpretability of medical image diagnostic analysis to some extent, the lack of objective quantitative information and the semantic ambiguity of qualitative conceptual factors can easily lead to ambiguity and diagnostic errors during consultations. Specifically, due to differences in individual experience, different doctors may have different criteria for judging whether the edge features of ultrasound nodules are clear or irregular, which can lead to deviations in their understanding of diagnostic criteria, resulting in serious problems of underdiagnosis or overdiagnosis. Conventional image processing quantitative methods are mainly used in unlabeled cases where feature information has prior guidance or significant rule constraints, and can extract relevant feature information such as intensity, texture, and shape using conventional medical image processing techniques. Especially in the quantitative analysis of ultrasound nodules, individual differences lead to variations in the image quality of ultrasound imaging, making it difficult to accurately quantify the morphological characteristic factors of nodules. Summary of the Invention

[0003] The purpose of this study is to propose an interpretable identification system based on multi-scale probability mapping and morphological feature quantification, including lesion contour extraction, uncertainty error estimation, morphological feature quantification, and interpretable lesion identification. This system effectively automates and standardizes the quantitative analysis of lesions in ultrasound images, further improving the accuracy and interpretability of lesion identification. It has significant application value for quantitative analysis of 2D ultrasound images and assisting clinicians in formulating accurate diagnosis and treatment plans.

[0004] To achieve the aforementioned technical objectives, the specific technical solution in this case is as follows.

[0005] Firstly, this case proposes an interpretable recognition system based on multi-scale probability mapping morphological feature quantization. The system includes an uncertainty error estimation module, a morphological feature quantization estimation module, and an interpretable recognition module. The uncertainty error estimation module is configured to learn the multi-scale feature distribution of lesion regions in ultrasound nodule images, obtain contour segmentation probability mappings of N lesion regions, and then obtain the lesion contour M. p Uncertainty error M of the profile e N is a set value; the morphological feature quantification estimation module is configured to be based on the lesion contour Mp and the uncertainty error M of the contour e obtain a quantitative representation of the nodule lesion shape feature; the explainable identification module is configured to first calculate the lesion region based on the uncertainty error M of the contour e calculate the lesion region, and then obtain image features of the lesion region, and visualize the lesion attribute based on the image features and the quantitative representation of the nodule lesion shape feature.

[0006] In an embodiment of the above technical solution, the uncertainty error estimation module includes a preset hybrid encoder and a multi-scale decoder; the preset hybrid encoder has a plurality of first feature encoding layers, encodes the original image and performs encoding feature splicing and fusion, and the multi-scale decoder has a plurality of second feature decoding layers, decodes and fuses the spliced and fused features to obtain fused features for predicting a probability map of the lesion region segmentation; wherein the number of second feature encoding layers is one less than the number of first feature decoding layers, and each second feature layer is connected to a first feature layer with a scale greater than or equal to the layer scale; a Monte Carlo pruning layer is arranged in the preset hybrid encoder and the multi-scale decoder to learn multi-scale feature distribution.

[0007] In an embodiment of the above technical solution, the parameters of the preset hybrid encoder are obtained through mask learning pre-training, and the steps include: denoting the preset hybrid encoder as E ct , constructing a hybrid encoder G ct with the same structure as the hybrid encoder E EM , and constructing a decoder G DM based on the Transformer at the same time; the original image is masked according to the set mask matrix and mask rate to obtain a convolution mask strategy and a mask image; in pre-training, the hybrid encoder G EM encodes the mask image based on the convolution mask strategy and splices and fuses the encoded features, the decoder G DM decodes the spliced and fused encoded features to reconstruct the original image, and an image structural similarity metric is used as a loss function to regularize and constrain the mean square error loss function; after pre-training, the parameters of the hybrid encoder G EM are used as the parameters of the hybrid encoder E ct .

[0008] In an embodiment of the above technical solution, the preset hybrid encoder has five first feature layers, which are divided into three convolution layers and two Transformer layers, and a Monte Carlo pruning layer is arranged at the third to fifth layers.

[0009] In an embodiment of the above technical solution, the multi-scale decoder is configured to have four second feature layers {U4, U3, U2, U1}, a Monte Carlo pruning layer is arranged in U4 and U3, and the four second feature layers {U4, U3, U2, U1} sequentially output decoding features as {F u4 , F u3 , F u2 , F u1}, wherein the feature F u2 is a decoding fusion of an up-sampling feature based on the feature F u3 and the feature F u4 , the feature F u1 is a decoding fusion of an up-sampling feature based on the feature F u2 and the features F u4 , F u3 .

[0010] In an embodiment of the above technical solution, the lesion contour M p and the uncertainty error of the contour M e , the obtaining step comprises: taking the i-th feedforward prediction ultrasound lesion suspicious area mapping probability as p mi , then the lesion contour is M p , N is the total number of times; taking the uncertainty cutoff value as Δp0, then the uncertainty error of the contour M e , δ 2 (*) represents a variance calculation function of the probability mapping.

[0011] In an embodiment of the above technical solution, the lesion morphology feature quantification comprises feature quantification of the scale, boundary, echo quality, composition and calcification of the lesion area.

[0012] In an embodiment of the above technical solution, the loss function is β is a balance factor, L is a mean square error loss function; μ is a pixel mean value of the ultrasound nodule image block, is a pixel mean value of the predicted ultrasound nodule image block; γ represents the covariance between image blocks, c b and c c respectively represent the brightness difference deviation and the contrast difference deviation.

[0013] In a second aspect, the present case proposes a computer readable storage medium, which stores a computer program capable of being loaded and executed by a processor to execute any of the above systems.

[0014] The beneficial technical effects of the case are: (1) providing an objective and quantitative approach to explain the decision-making process of the deep learning model through quantitative analysis of the morphological characteristics of the nodules, further increasing the credibility of the model, and prompting doctors to effectively use quantitative data in diagnostic analysis, improving the accuracy and reliability of ultrasound lesion identification, and providing data support with reference value for surgical resection or ablation treatment plan. (2) Through mask self-supervised learning, the heavy workload of manual annotation in deep learning is reduced, and the problem of difficult identification of nodule lesions caused by unclear image quality and noise interference due to individual differences in ultrasound nodule quantitative analysis can be overcome, realizing accurate representation of images and nodule morphological characteristics to meet the needs of clinical auxiliary diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0016] Figure 1 、 One A feature representation module structure based on hybrid mask self-encoding in an embodiment.

[0017] Figure 2 、 One An outline uncertainty error estimation schematic diagram using a hybrid encoder and a multi-scale decoder in an embodiment.

[0018] Figure 3 、 One A lesion outline feature quantization calculation schematic diagram using prior knowledge guidance in an embodiment.

[0019] Figure 4 、 One An interpretable identification process schematic diagram based on multi-scale probability mapping and morphological feature quantization in an embodiment.

[0020] Figure 5 、 One A thyroid ultrasound nodule lesion identification result and nodule outline morphological feature visualization effect schematic diagram in an embodiment.

[0021] Figure 6 、 One A thyroid ultrasound nodule lesion identification result and nodule internal morphological feature visualization effect schematic diagram in an embodiment. DETAILED DESCRIPTION

[0022] The case is based on medical image analysis process and ultrasound diagnosis guideline knowledge, using self-supervised learning, contour segmentation uncertainty evaluation, morphological feature quantification and other image processing technologies, to establish an interpretable recognition system of multi-scale probability mapping morphological feature quantification, which has important application value for quantitative analysis of 2D ultrasound images and assisting clinicians in formulating accurate diagnosis and treatment plans. The interpretability refers to the feature alignment and semantic fusion of lesion morphological feature quantization representation and image information, and the visual identification of lesion attributes. The lesion morphological feature quantization representation is the quantization representation of the features of the scale, boundary, echo, composition and calcification of the lesion area.

[0023] The technical solutions of the case will be described clearly and completely below in combination with the drawings. Obviously, the described embodiments are only part of the embodiments of the case, not all the embodiments. Based on the embodiments in the case, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0024] (I) Hybrid encoder G EM

[0025] In recent years, deep learning-based quantification methods are mainly applied to tasks where labeled, manual quantification estimation is difficult or inefficient. For some tasks where manual quantification estimation is difficult or inefficient, such as prediction of complex nonlinear relationships or trend prediction tasks, deep learning models can better handle these problems because they can automatically learn and extract complex correlation features, thereby improving the accuracy of continuous value prediction or estimation. The most common quantification prediction methods include convolutional neural networks (CNN), recurrent neural networks (RNN), or long short-term memory networks (LSTM), which are used to extract features from time series data and make predictions, or to extract key marker points, region of interest contours, and other information first. However, such methods require a large amount of labeled data to drive, and still cannot achieve accurate morphological feature quantization representation, thereby failing to meet the needs of clinical auxiliary diagnosis. The morphological feature quantization representation referred to here is the quantization representation of the scale, boundary, echo, composition, and calcification of the lesion area.

[0026] Based on this, the case is designed to reduce the amount of manual annotation and improve the accurate expression of ultrasound nodule morphological features. EM , combined with the decoder G DM , pre-trained using the mask learning prediction method, to realize a complete self-supervised learning framework, so as to realize the use of mask training without manual annotation, and to promote the model to realize the representation of the detailed features of the ultrasound nodule image.

[0027] In the hybrid encoder G EMThe decoder can have multiple feature coding layers. Each feature coding layer is guided by a convolutional masking strategy to encode and represent the features of the input ultrasound nodule image. The last feature coding layer concatenates and fuses the previously encoded features. Decoder G DM The coded features of the concatenated and fused image are decoded to predict the masked image patches, thus reconstructing the original image. During the training of the hybrid encoder, a brightness loss-based L... b Contrast loss L c Structural loss L s Construct a composite loss function to optimize the model.

[0028] See Figure 1 For example, three guided convolutional layers {C1, C2, C3} are used. Each of C1, C2, and C3 includes a patch-merged convolutional layer and a masked convolutional layer feature extraction module. The masked convolutional layer feature extraction module consists of a residual module merging Dropout layer and an Inplace-ABN activation layer. The convolutional kernels in the residual module perform convolution calculations according to the designed convolutional mask blocks to avoid leakage of feature information from the masked regions.

[0029] The mask blocks designed in convolutional layers C1, C2, and C3 are implemented by executing a convolutional masking strategy. See also Figure 1 (b) The right side shows the masked image. The original image is divided into small blocks, and random mask regions are set. The mask region has a corresponding mask matrix. The masking rate is m. For example, k = 14, m = 0.7. Based on the mask image, its serialized encoding M4 can be obtained, which is then converted into a convolutional mask matrix M3. Next, the convolutional mask is upsampled by a factor of two to obtain M2, and this operation is repeated to obtain M1. The upsampling factor can vary in different specific implementations. The sizes of M1, M2, and M3 are the same as the sizes of C1, C2, and C3. M3, M2, and M1 are three sets of mask convolutional regions of different scales, which can be denoted as follows: This is used to guide the computation of the convolutional layers. Specifically, the convolutional kernel of C1 is convolved with the mask blocks in M1, the convolutional kernel of C2 is convolved with the mask blocks in M2, and the convolutional kernel of C3 is convolved with the mask blocks in M3. The feature representations of the output image are denoted as F. c1 F c2 F c3 Next, the mask matrix of M4 is used. For convolutional features F c3 Serialization is performed to obtain the serialization encoding features of the input image.

[0030] Following the convolutional layer C3, a Transformer-based feature representation layer, denoted as T4, is constructed. The serialized encoded features are input into T4, and its corresponding output feature representation is F.t4 .

[0031] After T4, a layer of Transformer T5 is connected again, T5 uses attention mechanism to process the fusion features F c1 , F c2 , F c3 , F t4} obtained by the weighted and spliced fusion operation of {w1, w2, w3, w4} to obtain the fusion feature F t5 , as shown in formula (1):

[0032] F t5 = concat[w1F c1 , w2F c2 , w3F c3 , w4F t4 ] (1)

[0033] Where w1, w2, w3, w4 are attention weights.

[0034] The above five feature representation layers, namely {C1, C2, C3, T4, T5}, constitute a hybrid mask autoencoder feature representation module G EM , output features {F c1 , F c2 , F c3 , F t4 , F t5}. The mask matrix and mask rate established by the above process guide the image features to be serialized and expanded as input data of the Transformer layer, thereby realizing seamless connection of convolution and Transformer calculation.

[0035] The decoder G DM is established based on Transformer, which is asymmetric with the hybrid encoder G EM . The decoder and the hybrid encoder G EM constitute a hybrid autoencoder self-supervised learning model, which can reduce the workload of manual labeling and improve the accurate expression of ultrasonic lesion morphological features after training.

[0036] The decoder G DM uses F t5 and the convolution mask strategy to decode and reconstruct the image mask area This process can be represented as:

[0037]

[0038] Image structural similarity is used to regularize and constrain the mean square error loss function, thereby forming a composite loss function L co , as shown in formula (3), which optimizes the enhanced hybrid encoder GEM and decoder G DM characterization ability of the ultrasound nodule image.

[0039]

[0040] where β is a balance factor, L MSE is a mean square error loss function; μ is a pixel mean value of the ultrasound nodule image block, is a pixel mean value of the predicted ultrasound nodule image block; γ represents a covariance between image blocks, c b = (K1L) 2 and c c = (K2L) 2 respectively represent luminance difference deviation and contrast difference deviation, where L is a gray scale number 255, and K1<<1 and K2<<1. In an embodiment, K1≈0.01, K2≈0.03.

[0041] The composite loss function L co adjusts the punishment degree through the balance factor β, and optimizes the enhanced hybrid encoder G EM and the first decoder G DM characterization ability of the ultrasound nodule image.

[0042] The hybrid encoder G EM and the decoder G DM are pre-trained using mask learning, so as to realize a complete self-supervised learning framework, so as to mine the feature characterization information of the unlabeled ultrasound nodule image.

[0043] (ii) Uncertainty error estimation module

[0044] After the pre-training of the hybrid encoder G EM , the convolutional mask operation thereof is removed, and the hybrid coding structure of convolution and Transformer is reserved, so as to construct a hybrid encoder E ct for lesion segmentation, and parameters of the hybrid encoder G EM are used as parameters of the hybrid encoder E ct .

[0045] Referring to Figure 2 , the hybrid encoder E ct includes five first feature layers {C1', C2', C3', T4', T5}, including 3 convolutional layers and 2 Transformer layers, which respectively output side features {F c1 , F c2 , F c3 , F t4 , F t5}。C1′, C2′, C3′, T4′ correspond to C1, C2, C3, T4 in turn, and have the same size, C1′, C2′, C3′, T4′ reuse the parameters of C1, C2, C3, T4 (i.e. C1′ corresponds to C1, i = 1, 2, 3; T4′ corresponds to T4), and the convolution layers involved perform feature extraction on the complete image in a conventional convolution calculation manner.

[0046] On this basis, a multi-scale decoder D m is constructed. m The decoder D u4 includes four second feature layers {U4, U3, U2, U1}, and the number of second feature layers is one less than the number of first feature layers, thereby forming a symmetrical structure with the hybrid encoder. The scales of the second feature layers {U4, U3, U2, U1} correspond to the scales of {T4′, C3′, C2′, C1′}. Each second feature layer is connected to a first feature layer with a scale greater than or equal to that of the second feature layer, so as to reuse all intermediate features of different layers and scales to complete the fusion task. The four second feature layers {U4, U3, U2, U1} each output features {F u3 , F u2 , F u1}, and the multi-scale features are decoded and fused, i.e. {U1: U1+U3+U4, U2: U2+U4}, further enhancing the constraint of semantic information in D m on the prediction result, and thereby improving the sensitivity to the multi-scale features of the lesion. u2 , F u3 , F u4 , F u1 , F u2 , F c4 , F u3 .

[0047] The decoder D m and the encoder E ct constitute a backbone network module Useg for ultrasound nodule image lesion segmentation, which realizes preliminary segmentation of the nodule lesion in the ultrasound nodule image and obtains N lesion contour probability maps, where N is a set value.

[0048] An uncertainty error estimation module is formed by introducing a Monte Carlo pruning layer in the {C3′, T4′, T5, U4, U5} layers in the backbone network module Useg. The uncertainty error estimation module uses MC-Dropout operation to prompt the model to learn the feature distribution of the image to enhance the robustness of the model. The uncertainty error estimation module is based on the fused feature F u1to predict the probability map of lesion region segmentation, and then obtain the lesion contour M p and the uncertainty error of the contour M e .

[0049] The uncertainty error estimation module uses a 2D ultrasound label dataset to train the uncertainty error estimation network Mseg.

[0050] In an embodiment, the U4 structure is sequentially composed of a multi-scale feature merging layer, a multi-head self-attention layer, and an MC-Dropout layer. The U3 structure is sequentially composed of a Patch expansion up-convolution layer, a multi-scale feature merging layer, a multi-head self-attention layer, and an MC-Dropout layer (Monte Carlo pruning layer). The U2 and U1 structures are composed of a Patch expansion up-convolution layer, a multi-scale feature merging layer, and a multi-head self-attention layer. The pruning probability Δp of the Monte Carlo pruning layer is 0.3, so as to obtain the posterior probability distribution of the model D MC weight. m If the suspicious region mapping probability in the single feedforward predicted ultrasound nodule image lesion suspicious region mapping probability map is x is the input ultrasound image, is the feature encoding of the ultrasound image after embedding the Monte Carlo pruning layer in the feature layer [C3', T4', T5], is the feature decryption probability after embedding the Monte Carlo pruning layer in the feature layer.

[0051] The mean of the obtained N lesion contour probability maps is taken to obtain the lesion contour M p and the uncertainty error of the contour M e as shown in formulas (4) and (5).

[0052]

[0053]

[0054] wherein δ 2 (*) represents a variance calculation function of the probability map, Δp0 is an uncertainty truncation value, i.e., an uncertainty probability peak value. p mi is the i-th feedforward predicted ultrasound lesion suspicious region mapping probability. Formula (5) represents the uncertainty error of the contour under the condition that the uncertainty truncation value is 0.5.

[0055] (Three) Morphological feature quantification estimation module

[0056] A nodule morphological feature quantification estimation module based on clinical prior knowledge is established to realize the quantitative representation of the scale, boundary, echo quality, composition, and calcification of the lesion region. The feature quantification of the scale, boundary, echo quality, composition, and calcification of the lesion region is denoted as Qa , Q m , Q e , Q p , Q c The estimation method is as follows.

[0057] Nodule size quantification Q a , is to measure the size of the nodule lesion by fitting an ellipse using the contour point set P n = f n (M p ), Based on the contour point set P n , the nodule size quantification Q a is estimated as shown in equation (6).

[0058]

[0059] Where f area (*) is an area calculation function, W is the length of the lesion area, h is the width of the lesion area, d is the depth of the lesion area, R w / h is the aspect ratio of the lesion area, S o is the area of the lesion area. Figure 3 is a schematic diagram of prior knowledge guided morphological feature quantification for the lesion area, where the contour of the lesion area is represented as M p , Based on the contour information, the geometric center coordinates O, the area S, the width w, the height h and the depth d and other size feature information are calculated, further by fitting an ellipse using the contour point set to measure the shape information of the nodule lesion, and by calculating the area of the protruding area S un and the concave area S n outside the ellipse part through the boundary concave-convex ratio function f n (P u ). The nodule size quantification Q a , can be used as a nodule boundary smoothness evaluation factor information, for auxiliary identification of benign and malignant lesions.

[0060] Contour quantification Q m , is to measure the smoothness of the measured ultrasound lesion contour. Based on the obtained lesion contour M p and M e , the contour quantification Q m is estimated as shown in equation (7).

[0061]

[0062] Where M rp , r a , r b respectively represent the fitted ellipse point set, the short axis and the long axis of the ellipse, f ellispse (*) is an ellipse fitting function; S n, S u , S area , S rp , S n , S m , S lm , S cm , S cmi , S e , S p , S e , S ei , S ei , S e , S e.nene , S e.low , S e.equ , S e . high , S ei , S p , S c , S c1 , S c2 , S c3 . b , S cb , S w , S cw .

[0063] , S e , S p , S e , S ei , S ei , S e , S e.nene , S e.low , S e.equ , S e . high , S ei , S p , S c , S c1 , S c2 , S c3 . b , S cb , S w , S cw .

[0064]

[0065] , S ei , S ei , S e , S e.nene , S e.low , S e.equ , S e . high , S ei , S p , S c , S c1 , S c2 , S c3 . b , S cb , S w , S cw .

[0066] , S p , S c , S c1 , S c2 , S c3 . b , S cb , S w , S cw ., which represents the nodule component feature quantitative value.

[0067] Calcification feature quantitative Q c , by first establishing the contour to the geometric center of the four equal line, the nodule is divided into four annular regions, and on this basis, based on the plane coordinate system, it is divided into four quadrants, so as to quantify the calcification characteristic information. Based on the obtained lesion contour M p , the calcification feature quantitative Q c is estimated as shown in equation (9).

[0068]

[0069] Where t ci is the component of the gray image threshold t c , η i represents the calcification echo threshold proportion compared with the pixel value 255, ranging from 70% to 98%; S ci is the area of the calcification region and has the area function f area (*) and the threshold calcification region extraction function f echo (*) calculation, x represents the input image; r ci = ψ ratio (*) represents the ratio of the calcification area to the envelope area under the corresponding threshold, reflecting the dispersion degree of the calcification region; L ci = Φ pos (*) represents the orientation information of the calcification connected region, f contours (*) represents the contour point extraction of the calcification connected region, l lines represents the four equal lines based on the lesion contour, l lines = {l1, l2, l3, l4} and l4 is the contour line, D directions represents the four-quadrant orientation based on the geometric center of the lesion region in the plane coordinate, through which two information can obtain the orientation information of the calcification point in the lesion.

[0070] (Four) interpretable ultrasound lesion recognition module

[0071] In the interpretable ultrasound lesion recognition module, the quantitative information is aligned with the lesion region image features, and the semantic fusion is realized, so as to realize the effective explanation of the accurately recognized ultrasound nodule lesion.

[0072] Specifically, the quantitative information {Q a , Q m , Q e , Q p , Q c} obtained by using the above estimation method is used as an explicit feature. At the same time, the uncertainty error M eThe suspicious lesion area is calculated and identified, and the suspicious lesion area is denoted as X. ROI A classification network was constructed to extract image features of suspicious lesion areas, and these image features were used as implicit feature information F. x The cross-attention mechanism is used to input both explicit and implicit features into the lesion identification module R. bm This enables accurate identification of benign and malignant lesions, visualizes quantitative information, and improves the interpretability of the identification model in assisted diagnosis.

[0073] In conclusion, Figure 4 The diagram illustrates the interpretable recognition process utilizing multi-scale probability mapping and morphological feature quantization. Figure (a) shows the feature representation module of the hybrid mask autoencoder, which seamlessly connects the convolutional layer and the Transformer using a convolutional masking strategy to establish the feature representation module G of the hybrid mask autoencoder. EM And the structural similarity regularized composite loss function is used to optimize L co Asymmetric decoder G DM Ultrasound nodule image mask learning enables the model to more efficiently represent features of unlabeled ultrasound nodule image data. Figure (b) shows the uncertainty error estimation module (Mseg), which estimates the uncertainty error by using G... EM Remove the masking strategy in the code and build a hybrid encoder E. ct Furthermore, a symmetric decoder D is established using hierarchical connections and multi-scale feature fusion operations. m Furthermore, a contour uncertainty error estimation module, Mseg, is constructed by introducing Monte Carlo pruning into the high-dimensional semantic structure layer to capture the probability mapping and contour uncertainty error of the segmentation results of ultrasound nodule lesions. Figure (c) shows the nodule morphology feature quantification estimation and interpretability identification module. Based on the contour information and contour uncertainty error in the segmentation results, it establishes a prior knowledge-guided lesion morphology feature quantification module using prior knowledge combined with image processing. This module achieves a quantitative representation of the scale, boundary, echo, composition, and calcification of the lesion region, where the scale quantification can be seen in the boundary attributes. Furthermore, a multi-information fusion interpretability identification module is constructed based on this. Utilizing a cross-attention mechanism, explicit quantification information and implicit image feature information of suspicious lesion regions are aligned and fused, further improving the model's accuracy in identifying ultrasound nodule lesions and effectively enhancing its interpretability.

[0074] The terms "first" and "second" used above are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature.

[0075] (V) Application

[0076] The lesion explainability identification module is composed of the backbone network module Useg for ultrasound nodule image lesion segmentation, the contour uncertainty error estimation module, and the nodule morphological feature quantification estimation and explainability identification module, wherein the backbone network module Useg for ultrasound nodule image lesion segmentation is trained by ultrasound nodule images with specific contour labels.

[0077] The lesion explainability identification module is used to identify the thyroid ultrasound nodule lesion in explainability, and the contour morphological feature (nodule boundary quantification) visualization is seen in Figure 5 , wherein (a) is the ultrasound nodule identification result and contour prediction result visualization, (b) is the nodule identification probability mapping visualization, (c) is the contour uncertainty error visualization, and (d) is the nodule boundary quantification representation visualization; and the nodule internal morphological feature visualization is seen in Figure 6 , wherein (a) is the ultrasound nodule identification result and contour prediction result visualization, (b) is the calcification quantification visualization, (c) is the composition quantification visualization, and (d) is the echo quality quantification visualization. Through the description of the above embodiments, those skilled in the art can clearly understand that the system of the present disclosure can be realized by means of software and necessary general hardware, of course, it can also be realized by special hardware including special integrated circuits, special CPUs, special memories, special components, etc. Generally, functions completed by computer programs can be easily realized by corresponding hardware, and the specific hardware structure for realizing the same function can also be various, such as analog circuit, digital circuit or special circuit, etc. However, for the present disclosure, software program implementation is a better embodiment.

[0078] Although the embodiments of the present application are described above in combination with the drawings, the present application is not limited to the above specific embodiments and application fields, and the above specific embodiments are only illustrative and guiding, but not limiting. Those skilled in the art can make many forms under the guidance of the present disclosure and without departing from the scope protected by the claims of the present application, which all belong to the protection of the present application.

Claims

1. An interpretable recognition system based on multi-scale probability mapping morphological feature quantization, characterized in that: The system includes an uncertainty error estimation module, a morphological feature quantization estimation module, and an interpretability identification module; The uncertainty error estimation module is configured to learn the multi-scale feature distribution of the lesion region in the ultrasound nodule image, obtain the contour segmentation probability mapping map of N lesion regions, and then obtain the lesion contour. Uncertainty error of the profile N is a set value, and the lesion outline is... Uncertainty error of the profile The acquisition steps include: denoting the mapping probability of the i-th feedforward predicted ultrasound lesion suspicious area as... The outline of the lesion for: N is the total number of trials; the uncertainty cutoff value is denoted as The uncertainty error of the profile : ,in, The variance calculation function of the probability map is represented; the uncertainty error estimation module includes a preset hybrid encoder and a multi-scale decoder; the preset hybrid encoder has multiple first feature encoding layers, which encode the original image and perform feature splicing and fusion; the multi-scale decoder has multiple second feature decoding layers, which decode and fuse the spliced ​​and fused features to obtain fused features of the probability map used for predicting lesion region segmentation; wherein, the number of second feature encoding layers is one less than the number of first feature decoding layers, and each second feature layer is skipped to a first feature layer with a scale greater than or equal to that layer; Monte Carlo pruning layers are set in the preset hybrid encoder and the multi-scale decoder to learn the multi-scale feature distribution; The morphological feature quantification estimation module is configured based on the lesion contour. Uncertainty error of the profile The morphological features of nodular lesions are quantitatively represented, including the scale, boundary, echo quality, composition, and calcification features of the lesion region. The interpretability identification module is configured to first assess the uncertainty error of the contour. The lesion area is calculated to obtain the image features of the lesion area. Then, based on the image features and the morphological features of the nodule lesion, the lesion attributes are visualized.

2. The system according to claim 1, characterized in that, The parameters of the preset hybrid encoder are obtained through mask learning pre-training, and the steps include: The preset hybrid encoder is denoted as Building and Hybrid Encoders Hybrid encoders with the same structure At the same time, a Transformer-based decoder is built. ; The original image is masked according to the set mask matrix and mask rate to obtain the convolution masking strategy and mask image; In pre-training, the hybrid encoder Based on a convolutional masking strategy, the masked image is encoded and the encoded features are concatenated and fused. The decoder... The encoded features of the splicing and fusion are decoded to reconstruct the original image. The loss function used is the image structure similarity measure to regularize the constrained mean squared error loss function. After pre-training, the hybrid encoder The parameters are used as a hybrid encoder. The parameters.

3. The system according to claim 2, characterized in that, The preset hybrid encoder has five first feature layers, consisting of three convolutional layers and two Transformer layers, with Monte Carlo pruning layers set in layers 3 to 5.

4. The system according to claim 1, characterized in that, The multi-scale decoder is configured to have four second feature layers. ,exist The structure includes a Monte Carlo pruning layer and four second feature layers. The decoded features are output sequentially as follows , Among them, features For feature-based The upper sampling feature and feature Decoding fusion, features For feature-based The upper sampling feature and feature ,feature Decoding and fusion.

5. The system according to claim 2, characterized in that, The loss function is: in, As a balance factor, The mean squared error loss function; The pixel mean of the ultrasound nodule image patch. To predict the pixel mean of an ultrasound nodule image patch; This represents the covariance between image patches. and These represent the brightness difference deviation and the contrast difference deviation, respectively.

6. A computer-readable storage medium, characterized in that: The system stores a computer program that can be loaded by a processor and executed by the system as described in any one of claims 1 to 5.

Images