Domain Generalization Mitigates Scanner-Induced Domain Shift in Medical Imaging.
TL;DR
The results show that DG methods, particularly those that explicitly regularize the learning process, improve out-of-domain generalization, but do not fully close the gap with in-domain performance, underscore the necessity of incorporating DG strategies to develop clinically deployable AI models.
Abstract
Deep learning models for medical image analysis often fail in clinical deployment due to domain shift from varied acquisition hardware and protocols. We present a comprehensive evaluation of various domain generalization (DG) techniques to mitigate this performance degradation. We evaluate six DG algorithms against a baseline on two distinct tasks using large, multi-institutional datasets: grading prostate cancer aggressiveness from MRI using the ProstateNet dataset and assessing breast density from mammograms using the DMIST dataset, using a leave-one-domain-out protocol. Our results show that DG methods, particularly those that explicitly regularize the learning process, improve out-of-domain generalization, but do not fully close the gap with in-domain performance. On the ProstateNet dataset, the FISH algorithm achieved the highest average out-of-domain AUROC (0.678), a statistically significant improvement over the baseline (0.613). We observed similar trends on the DMIST dataset. These findings underscore the necessity of incorporating DG strategies to develop clinically deployable AI models.