AUTHOR=Aftab Rukhma , Yan Qiang , Zhao Juanjuan , Yong Gao , Huajie Yue , Urrehman Zia , Mohammad Khalid Faizi 

TITLE=Neighborhood attention transformer multiple instance learning for whole slide image classification

JOURNAL=Frontiers in Oncology

VOLUME=Volume 14 - 2024

YEAR=2024

URL=https://www.frontiersin.org/journals/oncology/articles/10.3389/fonc.2024.1389396

DOI=10.3389/fonc.2024.1389396

ISSN=2234-943X

ABSTRACT=Pathologists utilize whole slide images (WSIs) of tissue section biopsies to identify tumor cells and subtypes for diagnosing cancer. Deep learning-based models, particularly weakly supervised ones, have emerged as a viable option. These methods classify WSIs using image parts or tiles  along with attention ratings. However, due to the heterogeneous nature of tumors, these methods often overlook the possibility of false positives and negatives. Both cancerous and healthy cells can proliferate in patterns that extend beyond individual tiles, leading to errors at the tile level that result in inaccurate tumor-level classification. To address this limitation, we introduce NATMIL (Neighborhood Attention Transformer Multiple Instance Learning), an innovative deep learning approach that employs the Neighborhood Attention Transformer. NATMIL leverages dependencies among WSI tiles, incorporating contextual constraints as prior knowledge into multiple instance learning models. Our approach enhances the accuracy of tumor classification by considering the broader tissue context, thus reducing errors associated with isolated tile analysis. We conducted a quantitative analysis to evaluate NATMIL’s performance against other weakly supervised algorithms. When applied to subtyping non-small cell lung cancer (NSCLC) and  lymph node (LN) tumors, NATMIL demonstrated superior accuracy. Specifically, NATMIL achieved accuracy values of 89.6% on the Camelyon dataset and 88.1% on the TCGA-LUSC dataset, outperforming existing methods. These results underscore NATMIL’s potential as a robust tool for improving the precision of cancer diagnosis using WSIs.