SoftSeg: Advantages of soft versus binary training for image segmentation

•SoftSeg is a deep learning training approach yielding soft outputs.•SoftSeg overcomes the problem of partial volume and model confidence calibration.•Soft features include a soft ground truth, a ReLU activation, and a regression loss.•Better segmentation performance was validated on three open-source MRI datasets.•SoftSeg enables to encode uncertainty and inter-rater variability.