Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
The paper addresses weakly supervised training of 3D medical image classifiers using only volume-level labels, focusing on improving slice-level prediction accuracy. They identify that a simple center-focused baseline outperforms attention-based MIL methods in this setting. To address this, they introduce Normal Guidance, a regularization technique that encourages a bell-shaped attention distribution. Normal Guidance significantly improves slice-level localization for attention-based and transformer-based MIL methods across three large medical imaging datasets.
Attention mechanisms in medical image analysis are often outperformed by a simple center-focused baseline, but a novel "Normal Guidance" regularization can reverse this trend, unlocking significantly better slice-level localization.
We consider training classifiers for 3D medical images using only one binary label for the entire volume rather than a label for each 2D slice. In such weakly supervised settings, can we learn accurate classifiers for slice-level predictions? Attention-based multiple instance learning (MIL) can produce an attention score for every slice. Yet recent work demonstrates that a simple center-focused baseline that ignores image content can outperform attention-based and transformer-based MIL at slice-level classification of 3D brain scans. We show this baseline also outperforms existing MIL at slice-level classification of thoracic and abdominal CT scans. Motivated by this baseline, we propose Normal Guidance, a regularization technique that encourages the learned attention distribution to follow a bell-shaped curve. Across three medical imaging datasets totaling over 4 million 2D slices, we show our Normal Guidance enables attention-based and transformer-based MIL methods to deliver significantly better slice-level localization than the state-of-the-art while remaining competitive at whole-scan classification.