2020/11/17 by Charley Gros, Gros, Charley, Andréanne Lemay +3 · 2 citations
Neuroscience · Medicine · Biochemistry, Genetics and Molecular Biology · #Brain Tumor Detection and Classification #Radiomics and Machine Learning in Medical Imaging #MicroRNA in disease regulation
paper · pdf · doi:10.48550/arxiv.2011.09041
Most image segmentation algorithms are trained on binary masks formulated as\na classification task per pixel. However, in applications such as medical\nimaging, this "black-and-white" approach is too constraining because the\ncontrast between two tissues is often ill-defined, i.e., the voxels located on\nobjects' edges contain a mixture of tissues. Consequently, assigning a single\n"hard" label can result in a detrimental approximation. Instead, a soft\nprediction containing non-binary values would overcome that limitation. We\nintroduce SoftSeg, a deep learning training approach that takes advantage of\nsoft ground truth labels, and is not bound to binary predictions. SoftSeg aims\nat solving a regression instead of a classification problem. This is achieved\nby using (i) no binarization after preprocessing and data augmentation, (ii) a\nnormalized ReLU final activation layer (instead of sigmoid), and (iii) a\nregression loss function (instead of the traditional Dice loss). We assess the\nimpact of these three features on three open-source MRI segmentation datasets\nfrom the spinal cord gray matter, the multiple sclerosis brain lesion, and the\nmultimodal brain tumor segmentation challenges. Across multiple\ncross-validation iterations, SoftSeg outperformed the conventional approach,\nleading to an increase in Dice score of 2.0% on the gray matter dataset\n(p=0.001), 3.3% for the MS lesions, and 6.5% for the brain tumors. SoftSeg\nproduces consistent soft predictions at tissues' interfaces and shows an\nincreased sensitivity for small objects. The richness of soft labels could\nrepresent the inter-expert variability, the partial volume effect, and\ncomplement the model uncertainty estimation. The developed training pipeline\ncan easily be incorporated into most of the existing deep learning\narchitectures. It is already implemented in the freely-available deep learning\ntoolbox ivadomed (https://ivadomed.org).\n