Reproduce “Multi-Stage Pathological Image Classification using Semantic Segmentation” on OpenResearch

About this paper

Shusuke Takahama, Yusuke Kurose, Yusuke Mukuta, Hiroyuki Abe, Masashi Fukayama, Akihiko Yoshizawa, Masanobu Kitagawa, Tatsuya Harada

Researchers at The University of Tokyo and RIKEN developed a multi-stage approach for Whole Slide Image analysis, combining patch-based feature extraction with whole-slide semantic segmentation. Their method, including a memory-efficient end-to-end learning strategy, achieves superior diagnostic accuracy, reaching up to 99.34% PR-AUC on stomach biopsy datasets and consistently outperforming traditional patch-based methods on public datasets like Camelyon16 and Camelyon17.

Key insights

• Integrating local patch features with whole-slide global context is crucial for accurate WSI analysis, mimicking human pathologist workflow.
• Memory-efficient gradient retention allows for end-to-end training of multi-stage deep learning models on gigapixel images, overcoming GPU memory limitations.
• Combining a patch-based feature extractor with a U-Net for semantic segmentation yields smoother, more coherent prediction maps compared to purely patch-based methods.

Summary by alphaXiv.