アイテム詳細

要旨

Accurately segmenting MRI images is crucial for many cli-nical applications. However, manually segmenting imageswith accurate pixel precision is a tedious and time consu-ming task. In this paper we present a simple, yet effectivemethod to improve the efficiency of the image segmenta-tion process. We propose to transform the image annota-tion task into a binary choice task. We start by using classi-cal image processing algorithms with different parametervalues to generate multiple, different segmentation masksfor each input MRI image. Then, instead of segmenting thepixels of the images, the user only needs to decide whethera segmentation is acceptable or not. This method allowsus to efficiently obtain high quality segmentations with mi-nor human intervention. With the selected segmentations,we train a state-of-the-art neural network model. For theevaluation, we use a second MRI dataset (1.5T Dataset),acquired with a different protocol and containing annota-tions. We show that the trained network i) is able to au-tomatically segment cases where none of the classical me-thods obtain a high quality result ; ii) generalizes to thesecond MRI dataset, which was acquired with a differentprotocol and was never seen at training time ; and iii) en-ables detection of miss-annotations in this second dataset.Quantitatively, the trained network obtains very good re-sults : DICE score - mean 0.98, median 0.99- and Haus-dorff distance (in pixels) - mean 4.7, median 2.0-.