Research On Medical Image Classification Algorithms Based On Few-Shot Learning

Few-shot learning are algorithms based on deep learning frameworks that aim to achieve high generalization performance using just few reference samples provided in categories never met in training procedure,which are well behaved in natural image classification tasks.However,similar algorithms are insufficient for medical image datasets under cross domain settings by the reason of inadequate feature-extracting capability for mainstream frameworks and thus leading to poor generalization performance.This thesis focuses on how to improve average generalization performance of few-shot classification framework on medical image datasets by introducing more powerful backbones and designing more effective learning strategies,aiming to improve the feature extraction capability of deep learning frameworks with similar model size.The framework is optimized from three aspects,such as structure and weight fine-tuning strategy of feature extraction network,stage division optimization for few-shot learning process and a new learning strategy.Experiments show that the optimized framework can extract enhanced features for decision-making,thus helping boost performance of framework.1.A lightweight Vision Transformer(ViT)was introduced into the few-shot medical image classification framework based on CNNs.The information of different image patches provided by reference samples was fully utilized.At the same time,in the pre-training stage,the number of categories of training dataset was enlarged.In the few-shot learning stage,partial weight fine-tuning strategy was adopted to update parameters,which ensured better classification accuracy after limited learning epochs.2.The stage division of the learning process was optimized,extending the two-stage training into three-stage training,thus forming a learning process with progressive difficulties from classification tasks with fixed classes,to contrastive learning tasks for dynamic classes,and then to fewshot classification tasks for dynamic classes.3.A joint goal learning strategy was employed by combining distillation loss and classification loss.In addition,a teacher-studentnetwork guided task difficulty rating algorithm was investigated to filter tasks before training.In order to verify cross-organ generalization performance of framework,three pathological image datasets of colon,skin and blood cells were used in this thesis to conduct experiments in which average classification accuracy was evaluated.