Using Off-Target-Disease Data to Improve Target-Disease Lesion Segmentation

Does increasing the training data in deep learning always improve the accuracy of lesion segmentation on brain MRIs? To answer this question, we conduct experiments of small-diffuse lesion segmentation on 8 public brain MRI datasets with 3 different diseases, including Multiple Sclerosis (MS), White Matter Hyperintensities (WMH), and Chronic Stroke (CS). These diseases all have small lesions (lesions occupying < 2% of the whole brain MRI on average) that are often diffuse (injuring multiple brain regions), posing considerable challenges for lesion segmentation using deep learning. Our experimental results show that adding more data from other off-target diseases often led to a surprising decrease in accuracy on the target disease using a Naïve training strategy. To address this, we propose a disease-switch training strategy to learn the knowledge from off-target diseases to the target disease. The accuracies on the three diseases show that the proposed disease-switch training strategy consistently improves the accuracy of small-diffuse lesion segmentation on the three diseases.