Prediction of IDH Status Using Hierarchical Attention-Based Deep 3D Multiple Instance Learning

Motivation: Accurate prediction of isocitrate dehydrogenase (IDH) mutations from multimodal MRI remains challenging due to tumor heterogeneity. Goal(s): We proposed a Hierarchical Attention Based Multi Instance Learning (HAB-MIL) framework for preoperative IDH prediction. Approach: This method utilizes features associated with IDH mutations and incorporates positional encoding to enhance pixel position cues. Additionally, it employs attention-based gated pooling on 3D instances to improve the understanding of labels. Results: Our method achieved an AUC of 91.1% and accuracy of 93.7% on the TCIA dataset, outperforming state-of-the-art techniques. Finally, Grad-CAM was employed to visualize the results of the model. Impact: The incorporation of a dynamic attention mechanism in HAB-MIL effectively explores tumor-related features for IDH prediction, while also fully leveraging tumor positional information. This enhancement improves the model's interpretability, providing more valuable support for clinical diagnosis.