Multi-stage knowledge distillation with layer fusion-based deep learning approach for skin cancer classification

基于层融合的多阶段知识蒸馏深度学习方法用于皮肤癌分类

阅读:1

Abstract

Skin cancer is one of the most common types of cancer globally, caused by prolonged exposure to the sun's UV rays. Despite recent developments in research, early diagnosis, prevention, and treatment, skin cancer remains a significant health concern. This study proposes a multi-stage knowledge distillation-based deep learning technique with a layer fusion strategy to classify different types of skin lesion cells using the HAM10000 dataset. Augmentation and basic preprocessing steps have been applied to the HAM10000 dataset to enhance robustness. The applied multi-stage knowledge distillation incorporates intermediate features by measuring several loss values and coefficients to balance the corresponding losses. The proposed ViT and ConvNeXT-integrated teacher model leverages hybrid architectures derived from baseline models, combining convolutional feature extraction with transformer-based attention mechanisms. The distilled model, built using CNN and EfficientNet, achieved significant performance improvements over the baseline. The optimized model achieved an accuracy of 95.88%, F1 and AUC scores of 95.91% and 99.02%, respectively. Multi-stage knowledge distillation with intermediate exits and layer fusion improved model accuracy and performance metrics with the lowest training and inference times of 61 and 15 ms/step, respectively. Post-training quantization was applied to reduce the parameters and size of the distilled model. Multiple XAI techniques-Grad-CAM, Score-CAM, and LIME-have been explored to enhance the interpretability of the applied multi-stage knowledge distillation model. The implementation codes with the quantized and distilled model are available in the following repository: https://github.com/codewith-pavel/Optimizations .

特别声明

1、本页面内容包含部分的内容是基于公开信息的合理引用;引用内容仅为补充信息,不代表本站立场。

2、若认为本页面引用内容涉及侵权,请及时与本站联系,我们将第一时间处理。

3、其他媒体/个人如需使用本页面原创内容,需注明“来源:[生知库]”并获得授权;使用引用内容的,需自行联系原作者获得许可。

4、投稿及合作请联系:info@biocloudy.com。