SMILE: Sparse-Attention based Multiple Instance Contrastive Learning for Glioma Sub-Type Classification Using Pathological Images

Gliomas are the most prevalent malignant brain tumor in adults and can be classified into four typical sub-types based on histological features. Histological diagnosis by pathologists via microscopic visual inspection of pathological slides has been the gold standard for glioma grading, especially hematoxylin and eosin (H&E) sections. However, due to spatial heterogeneity and complex tumor micro-environment, it is difficult and time-consuming for pathologists to differentiate glioma sub-types. In this paper, we propose a Sparse-attention based Multiple Instance contrastive LEarning (SMILE) method for glioma sub-type classification. First, we use contrastive learning to extract meaningful representations from pathological images. Second, we propose the sparse-attention multiple instance learning aggregator to get sparse instance representations in a bag for label prediction. We validate the proposed SMILE method using a glioma dataset from The Cancer Genome Atlas (TCGA). Experimental results show superior performance of our method over competing ones. Ablation study further demonstrates the effectiveness of our design of SMILE.