GradMask: Reduce Overfitting by Regularizing Saliency
This addresses overfitting in medical imaging for clinicians by reducing reliance on non-tumor features, though it is incremental as it builds on existing regularization techniques.
The paper tackles overfitting in medical imaging by introducing GradMask, a regularization method that penalizes saliency maps inconsistent with lesion segmentation, resulting in a 1-3% improvement in test accuracy compared to baseline.
With too few samples or too many model parameters, overfitting can inhibit the ability to generalise predictions to new data. Within medical imaging, this can occur when features are incorrectly assigned importance such as distinct hospital specific artifacts, leading to poor performance on a new dataset from a different institution without those features, which is undesirable. Most regularization methods do not explicitly penalize the incorrect association of these features to the target class and hence fail to address this issue. We propose a regularization method, GradMask, which penalizes saliency maps inferred from the classifier gradients when they are not consistent with the lesion segmentation. This prevents non-tumor related features to contribute to the classification of unhealthy samples. We demonstrate that this method can improve test accuracy between 1-3% compared to the baseline without GradMask, showing that it has an impact on reducing overfitting.