Weakly-supervised Kidney Tumor Classification from CT Scans with Multi-Instance Learning and Anatomical Filtering

Developing deep learning models for CT scan analysis is typically bottlenecked by the high cost of pixel-level annotations. This paper investigates whether a weakly supervised pipeline—using only scan-level labels—can effectively classify kidney tumors while maintaining robustness across different clinical datasets. The authors specifically address the challenges of low supervision ratios and the computational burden of processing large 3D volumes.

The authors propose a two-part pipeline. First, they introduce Compass, a self-supervised body part regression model that maps CT slices to a continuous anatomical coordinate system. This allows the system to automatically crop scans to pathology-relevant regions without requiring manual segmentation masks. Second, they evaluate two Multi-Instance Learning (MIL) frameworks: Attention-based MIL (ABMIL), an embedding-based approach, and FocusMIL, an instance-based approach. These models treat CT volumes as a collection of 2D slices, allowing for the use of efficient 2D architectures pretrained on ImageNet.

The study finds that the Compass filtering module significantly reduces input data size (by 26% to 68%) while retaining critical tumor-containing slices. While both MIL architectures achieve strong performance on the internal test set (F1 = 0.83), they differ in their generalization behavior. Embedding-based ABMIL is prone to exploiting non-causal features, leading to poor performance on out-of-distribution data unless constrained by Compass filtering. In contrast, instance-based FocusMIL shows higher inherent robustness to distribution shifts. The combination of Compass filtering and MIL effectively prevents the model collapse often seen in 3D baseline models trained on limited data.

This work provides a scalable, annotation-efficient framework for medical image classification. By demonstrating that anatomical filtering can replace expensive segmentation masks, the authors offer a path toward deploying diagnostic models in clinical settings where pixel-level labels are unavailable. The results also highlight the importance of model architecture choice in ensuring that AI systems learn clinically relevant features rather than spurious correlations associated with specific imaging hardware.