关键词: 小样本学习, 图像分类, 皮肤病变, 空间变换网络, 最邻近质心

Abstract: Deep learning-based image classification methods typically require a large amount of labeled data. However, in the medical field, especially in the classification task of skin lesions, collecting a vast amount of image data faces numerous challenges. To accurately classify few-shot skin diseases, a few-shot classification model based on spatial transformation network and feature distribution calibration is proposed. First, transfer learning and meta-learning are integrated to address the overfitting issues present in cross-domain few shot transfer. Then, a rotation angle prediction task is inserted before the pre-training classification task to better adapt the model to the high complexity of medical image data. Subsequently, after downsampling the images, a spatial transformer network is introduced to explicitly perform affine transformations on the input images, enhancing feature extraction and recognition capabilities. Finally, feature distribution calibration is used to constrain new class features, and the nearest centroid algorithm is introduced for classification decisions, significantly reducing algorithm complexity while improving classification accuracy. Conduct experiments on the ISIC2018 skin image dataset: Compared to the current mainstream few-shot model Meta-Baseline, The proposed model has achieved an average accuracy improvement of 12.20 and 11.40 percentage points in 2-way and 3-way classification tasks, respectively; compared to MetaMed, The average accuracy has improved by 7.00 and 10.23 percentage points in the 2-way 3-shot and 3-way 3-shot classification tasks, respectively. The experimental results show that the proposed model effectively improves the classification accuracy of few-shot skin diseases, better assisting doctors in enhancing clinical diagnosis accuracy.

Key words: few-shot learning, image classification, skin lesions, Spatial Transformer Networks &#40, STN&#41