SoftSeg: Advantages of soft versus binary training for image segmentation

Most image segmentation algorithms are trained on binary masks formulated as a classification task per pixel. However, in applications such medical imaging, this black-and-white approach is too constraining because the contrast between two tissues often ill-defined, i.e., voxels located objects' edges contain mixture of (a partial volume effect). Consequently, assigning single hard label can result detrimental approximation. Instead, soft prediction containing non-binary values would overcome that limitation. In study, we introduce SoftSeg, deep learning training takes advantage ground truth labels, and not bound to predictions. SoftSeg aims at solving regression instead problem. This achieved by using (i) no binarization after preprocessing data augmentation, (ii) normalized ReLU final activation layer (instead sigmoid), (iii) loss function traditional Dice loss). We assess impact these three features open-source MRI datasets from spinal cord gray matter, multiple sclerosis brain lesion, multimodal tumor challenges. Across random dataset splittings, outperformed conventional approach, leading an increase score 2.0% matter (p=0.001), 3.3% for lesions, 6.5% tumors. produces consistent predictions tissues' interfaces shows increased sensitivity small objects (e.g., lesions). The richness labels could represent inter-expert variability, effect, complement model uncertainty estimation, which typically unclear with developed pipeline easily be incorporated into most existing architectures. implemented freely-available toolbox ivadomed (https://ivadomed.org).