Resolution-Based Distillation for Efficient Histology Image Classification.

摘要: Developing deep learning models to analyze histology images has been computationally challenging, as the massive size of causes excessive strain on all parts computing pipeline. This paper proposes a novel learning-based methodology for improving computational efficiency image classification. The proposed approach is robust when used with that have reduced input resolution and can be trained effectively limited labeled data. Pre-trained original high-resolution (HR) images, our method uses knowledge distillation (KD) transfer learned from teacher model student same at much lower resolution. To address lack large-scale datasets, we perform KD in self-supervised manner. We evaluate two datasets associated celiac disease (CD) lung adenocarcinoma (LUAD). Our results show combination self-supervision allows approach, some cases, surpass classification accuracy teacher, while being more efficient. Additionally, observe an increase performance unlabeled dataset increases, indicating there potential scale further. For CD data, outperforms HR model, needing 4 times fewer computations. LUAD 1.25x magnification are within 3% 10x magnification, 64 cost reduction. Moreover, outcomes benefit scaling use 0.625x using data improves by 4% over baseline. Thus, improve feasibility solutions digital pathology standard hardware.