AUTHOR=Wang Chunbao , Du Xianglong , Yan Xiaoyu , Teng Xiali , Wang Xiaolin , Yang Zhe , Chang Hongyun , Fan Yangyang , Ran Caihong , Lian Jie , Li Chen , Li Hansheng , Cui Lei , Jiang Yina 

TITLE=Weakly supervised learning in thymoma histopathology classification: an interpretable approach

JOURNAL=Frontiers in Medicine

VOLUME=Volume 11 - 2024

YEAR=2024

URL=https://www.frontiersin.org/journals/medicine/articles/10.3389/fmed.2024.1501875

DOI=10.3389/fmed.2024.1501875

ISSN=2296-858X

ABSTRACT=Thymoma classification is a challenging task due to its complex and diverse morphology. This study presents a novel AI-assisted diagnostic model that combines weakly supervised learning with a divide-and-conquer multi-instance learning (MIL) approach to enhance classification accuracy and interpretability. Using 222 thymoma slides, the model achieved a classification AUC of 0.9172 by simplifying the five-class task into binary and ternary steps. A key feature of this model is its attention-based mechanism, which generates interpretable heatmaps that align with clinically validated morphological differences between thymoma subtypes. This interpretability allows pathologists to visually confirm the AI's decisions, improving diagnostic reliability. By embedding domain-specific pathological knowledge into the interpretability framework, the study offers a new paradigm for making weakly supervised algorithms more transparent and applicable in clinical settings. This approach not only reduces the diagnostic burden on pathologists but also has the potential to improve patient outcomes.