CN-121544970-B - Medical image pathology detection model training method based on few sample learning

Abstract

The invention belongs to the field of image processing and discloses a medical image pathology detection model training method based on less sample learning, which comprises the following steps of S10, adopting a meta-learning paradigm, constructing less sample training fragments by random sampling from a training data set containing a plurality of category label images, simulating a real less sample classification scene by each fragment, comprising an example set and a target set, S20, constructing a shared feature encoder, processing the example set and the target set in each fragment, S30, generating a structured category anchor point based on a graph neural network for each category in the current fragment example set, carrying out sample weight self-adaptive distribution of the example set based on a meta-attention mechanism, S40, introducing a dynamic memory library and an uncertainty perception mechanism, carrying out category anchor point uncertainty quantification and history knowledge integration, and S50, classifying loss calculation and model updating to obtain a medical image pathology detection model. The invention realizes the accurate identification of pathological features of the medical image under the condition of few samples.

Inventors

• JIANG NING
• KUANG XIAOQI
• WANG LIANGJIE
• YU WENXIN
• HE GANG
• ZHANG ZHIQIANG

Assignees

• 西南科技大学

Dates

Publication Date

20260508

Application Date

20260120

Claims (5)

1. 1. A medical image pathology detection model training method based on less sample learning is characterized by comprising the following steps: s10, constructing a large number of training fragments with few samples from a dataset containing multi-category annotation images by adopting a meta-learning paradigm, wherein each fragment simulates a real classifying scene with few samples and comprises an example set and a target set; S20, constructing a shared depth convolution neural network as a feature encoder for uniformly extracting the features of the medical images of the example set and the target set in each segment; S30, generating a structured class anchor point by using a graph neural network according to the example set of each class of the current segment, and carrying out weight self-adaptive distribution on the sample of the example set based on a meta-attention mechanism so as to obtain the final representation of each class anchor point; the generation of the structured category anchor point comprises the following steps: S311, regarding feature vectors of all sample sets of the same class as graph nodes, constructing edges according to cosine similarity or k neighbor relations among the nodes, and forming a relation graph of the current class sample; S312, inputting the relation graph into a shallow graph convolution network, and aggregating neighbor node information of each node by means of a graph convolution message transmission mechanism to update self characteristics; s313, performing average pooling operation on all updated node characteristics, wherein the obtained vector is used as a structural category anchor point of the category; An example set sample weight adaptive allocation based on meta-attention mechanisms, comprising the steps of: s321, inputting sample characteristics of an example set into a meta-attention module formed by a lightweight feedforward neural network, and outputting attention weight of each sample by the meta-attention module by integrating the characteristics of the sample and intra-class relations; S322, weighting average is carried out on sample characteristics of the example set by using the obtained weight values, a weighted category anchor point is obtained, and the weighted category anchor point and the structured category anchor point are integrated for fine adjustment; S40, introducing a dynamic memory library and an uncertainty perception mechanism, quantifying the uncertainty of the category anchor points, and integrating with historical knowledge to obtain enhanced category anchor points; quantification and knowledge integration of category anchor uncertainty includes: calculating the variance of cosine similarity between each class of anchor points and the sample characteristics of the corresponding example set, and taking the variance as the uncertainty score of the current class of anchor points; Retrieving a plurality of historical category anchors most relevant to the current category anchor from a dynamic memory based on cosine similarity; dynamically adjusting the weight according to the uncertainty score, and weighting and integrating the current category anchor point and the retrieved historical category anchor point to form an enhanced category anchor point; s50, calculating classification loss based on cosine similarity between the target set sample and the enhanced category anchor points, and updating model parameters to obtain a final medical image pathology detection model.
2. 2. The training method of a medical image pathology detection model based on less sample learning according to claim 1, wherein in the step S10, an example set of each segment samples a image classes, and each class samples B random labeling samples; the target set consists of several unlabeled samples from the same a image categories for evaluating the classification performance of the model on the current segment.
3. 3. The training method of a medical image pathology detection model based on less sample learning of claim 1, wherein the shared feature encoder maps the input raw image pixel data to a low-dimensional embedding space, generating a corresponding feature vector representation.
4. 4. The training method of a medical image pathology detection model based on less sample learning according to claim 1, wherein in step S50, the cosine similarity between the sample characteristics of the target set and each enhanced class anchor point is calculated, converted into probability distribution by Softmax function, and the loss value is calculated by using cross entropy loss function.
5. 5. The training method of the medical image pathology detection model based on the less sample learning according to claim 1, wherein an error feedback mechanism based on an SGD optimization strategy is adopted to drive a feature extractor, a graph neural network and a meta-attention component to realize end-to-end parameter adaptation.

Description

Medical image pathology detection model training method based on few sample learning Technical Field The invention belongs to the technical field of image recognition, and particularly relates to a medical image pathology detection model training method based on less sample learning. Background In the field of medical image analysis, few sample learning is becoming a research hotspot gradually, and especially in scenes such as disease feature recognition and pathology detection, the traditional deep learning model is difficult to directly apply due to the fact that professional labeling data are scarce and the acquisition cost is high. The few sample medical image classification aims at training a model with a very small number of labeled samples, enabling it to accurately identify new classes of medical images, such as histopathological sections, dermatological images or optical coherence tomography images, etc. In order to overcome the bottleneck of scarce annotation data, researchers have proposed various technical ideas. The pre-training model based on transfer learning and the prototype network based on metric learning are two representative technical routes, and the training method is characterized. For a pre-training model scheme based on transfer learning, the deficiency is mainly reflected by the superposition effect of domain difference and data scarcity. First, this approach faces a serious information bottleneck. The feature space learned by the pre-training model on the natural image has essential differences with the semantic and texture features of the medical image, for example, the object outline and the color distribution in the natural image do not directly correspond to the cell structure and the lesion region features in the medical image. The domain difference causes that a large amount of labeling data is needed to complete characteristic adaptation of the model in a fine adjustment stage, but the labeling data is absolutely scarce under the condition of few samples, and complex characteristic mapping learning cannot be supported, so that the model is difficult to capture essential characteristics of the medical field, and the generalization capability is greatly reduced. Secondly, the method has a disadvantage in discriminant modeling between classes. The feature space initialized by the pre-training model is more suitable for distinguishing natural objects with obvious category differences, and medical image categories tend to be highly similar in vision, such as cancer cells of different subtypes or early lesions and normal tissues, and the subtle differences of the medical image categories need to be distinguished by highly specific feature representations. The migration learning scheme lacks an optimization mechanism aiming at fine granularity classification of medical images, so that classification boundaries are fuzzy, and diagnosis accuracy is affected. For prototype network schemes based on metric learning, the defects root in inherent limitations of class anchor construction and metric mechanisms, and the training process cannot effectively solve the problems. The first problem is that class anchor characterization is highly sensitive to noise and outliers. Class anchor calculation employs a simple arithmetic averaging method, which means that each sample in the example set contributes equally to the class anchor, including imaging artifacts, labeling errors, or atypical cases that may be present. In medical images, noise and outliers are common, such as motion artifacts in CT scanning or uneven staining in pathological sections, and a single outlier sample can significantly distort the position of a class anchor vector, destroying its representativeness. Second, the scheme is equally weak in discriminant modeling between classes. The classification decision depends only on the relative distance between the query sample and the class anchor point, lacking explicit optimization of the embedded spatial global structure. When different classes of medical images are visually highly similar, such as different stages of diabetic retinopathy, simple distance measures are difficult to learn to have a highly discriminatory decision boundary. In addition, the prototype network is designed to be isolated in segments, each segment is independently processed, knowledge in historical segments cannot be accumulated and multiplexed in the training process, the problem of information bottleneck caused by absolute scarcity of samples is further aggravated, and the applicability of the model in a continuous learning scene is limited. Disclosure of Invention In order to solve the problems, the invention provides a medical image pathology detection model training method based on less sample learning, which systematically integrates graph structural relation modeling, meta-attention driven weight distribution and cross-segment knowledge integration mechanisms into a unified meta-learning training