Loading...

Classification and Localization of Cancer in Histology Images Using Weak Labels

Jafarinia, Hossein | 2024

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 57897 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Rohban, Mohammad Hossein
  7. Abstract:
  8. Whole Slide Image (WSI) classification in digital pathology presents significant computational challenges due to the scale and complexity of the data. Traditional approaches often rely heavily on self-supervised learning (SSL), which demands extensive training periods and computational resources. Moreover, performance can suffer from domain shifts when transitioning from natural images to WSIs. We introduce Snuffy, a novel framework designed to address these challenges effectively. Snuffy utilizes a novel sparse transformer utilizing innovative sparse multi-head class-centered random global attention mechanism tailored specifically for pathology. This approach mitigates performance loss associated with limited pre-training and enables continual few-shot pre-training as a competitive option and utilizes the intuition that pathologists use about the tissue micro-environment, which enables it to model exceptionally difficult situations like cancer metastasis. We demonstrate Snuffy’s effectiveness on the CAMELYON16 and TCGA Lung cancer datasets, where it achieves superior performance in WSI and patch-level classifications by 0.132 improvement on WSI AUC while keeping the highest patch AUC in exceptionally difficult dataset of CAMELYON16, and Best or near best performance on TCGA Lung Cancer dataset. Also, Snuffy’s performance on other datasets like Musk1, Musk2, Elephant, and CIFAR-100 demonstrates its capability as a general framework. By integrating these advancements, Snuffy marks a significant progression in Multiple Instance Learning for digital pathology, delivering robust and precise results across a range of datasets. Additionally, we conduct ablation studies on every aspect of Snuffy to enhance the scientific community’s understanding of Snuffy and facilitate its more confident utilization in their practices.
  9. Keywords:
  10. Self-Supervised Learning ; Whole Slide Image (WSI) ; Multiple Instance Learning (MIL) ; Sparse Transformer ; Images Classification ; Cancer Diagnosis

 Digital Object List

 Bookmark

...see more