Observer variation-aware medical image segmentation by combining deep learning and surrogate-assisted genetic algorithms

There has recently been great progress in automatic segmentation of medical\nimages with deep learning algorithms. In most works observer variation is\nacknowledged to be a problem as it makes training data heterogeneous but so far\nno attempts have been made to explicitly capture this variation. Here, we\npropose an approach capable of mimicking different styles of segmentation,\nwhich potentially can improve quality and clinical acceptance of automatic\nsegmentation methods. In this work, instead of training one neural network on\nall available data, we train several neural networks on subgroups of data\nbelonging to different segmentation variations separately. Because a priori it\nmay be unclear what styles of segmentation exist in the data and because\ndifferent styles do not necessarily map one-on-one to different observers, the\nsubgroups should be automatically determined. We achieve this by searching for\nthe best data partition with a genetic algorithm. Therefore, each network can\nlearn a specific style of segmentation from grouped training data. We provide\nproof of principle results for open-sourced prostate segmentation MRI data with\nsimulated observer variations. Our approach provides an improvement of up to\n23% (depending on simulated variations) in terms of Dice and surface Dice\ncoefficients compared to one network trained on all data.\n