Multi-diagnosis multi-instance learning for auxiliary gene mutation diagnosis in whole slide images

Multi-Instance Learning (MIL) is widely applied in the representation of high-resolution whole slide images (WSIs) in pathology. However, there are still challenges in associating WSIs with genetic representations. The connection between the types of genetic mutations in WSI and their histological images is not intuitive, making it difficult for traditional models to effectively capture the key feature representations of genetic mutations.

This study proposes a novel cancer-gene multi-diagnosis multi-instance learning approach for WSI-level genetic mutation diagnosis. This method integrates cancer diagnostic data with pixel-level annotations and genetic mutation diagnostic data with WSI-level annotations as multi-diagnosis data inputs, systematically learning the representations of genetic mutations. By leveraging the high correlation between cancerous regions and genetic mutation areas, we filter out representative instances containing lesions and delve into the morphological and histological feature representations of genetic mutations through a focused sampling strategy.

Our proposed framework demonstrates advanced performance on gene mutation datasets of lung cancer, bladder cancer and breast cancer. Moreover, this method generates a series of visualizations that serve as interpretability tools reflecting model attention.

Our framework achieves deep coupling of cross-diagnostic pathological representations by deconstructing latent correlation maps between cancer regions and gene mutation regions in WSIs. Our method can, to some extent, provide decision support for pathologists' clinical diagnosis, and is potentially useful for triage or prioritization.