The article presents a new approach to dealing with computational limitations in developing computer vision solutions for gigapixel images in digital pathology. The authors propose a method of knowledge distillation that incorporates attention maps during training to enhance model performance at reduced image resolutions. The focus is on transferring knowledge about the most discriminative image regions, which may be lost when the resolution is decreased. The proposed approach has demonstrated substantial improvements in model performance across different image resolutions compared to previous methods.
Publication date: 12 Jan 2024
Project Page: https://github.com/cvblab/kd_resolution
Paper: https://arxiv.org/pdf/2401.06010
