Medical image segmentation automatic quality control: A multi-dimensional approach

• Published in: Medical image analysis Vol. 74; p. 102213
• Main Authors: Fournel, Joris, Bartoli, Axel, Bendahan, David, Guye, Maxime, Bernard, Monique, Rauseo, Elisa, Khanji, Mohammed Y., Petersen, Steffen E., Jacquier, Alexis, Ghattas, Badih
• Format: Journal Article
• Language: English
• Published: Netherlands Elsevier B.V 01.12.2021; Elsevier BV; Elsevier
• Subjects: Artificial Intelligence; Automatic control; Cardiology and cardiovascular system; CMR Image segmentation; Computer Science; Deep learning; Evaluation; Heart - diagnostic imaging; Heart Ventricles; Human health and pathology; Humans; Image enhancement; Image processing; Image Processing, Computer-Assisted; Image quality; Image segmentation; Imaging, Three-Dimensional; Life Sciences; Medical image segmentation automatic quality control; Medical imaging; Multi-dimensional quality control; Muscles; Myocardium; Quality control; Statistics; Ventricle; Deep learning; Medical image segmentation automatic quality control; CMR Image segmentation; 41A10; 65D05; 65D17; 41A05; Multi-dimensional quality control
• ISSN: 1361-8415; 1361-8423; 1361-8423
• DOI: 10.1016/j.media.2021.102213

•A new CNN-based segmentation quality control general approach, implementable for any MRI or computed tomography (CT) segmentation task, predicting both the 3D DSC and the 2D DSC.•A mathematical solution allowing to get 3D DSC predictions from 2D DSC and mean volume similarity fraction (MVSF) predictions.•A localization of segmentation errors at a 2D/slice level with a possibility to correct the effect of segmentation errors on clinical measurements.•A significant improvement of results obtained from state of the art approaches (for 3D DSC predictions) on CMR data, including an evaluation on a real-world application. [Display omitted] In clinical applications, using erroneous segmentations of medical images can have dramatic consequences. Current approaches dedicated to medical image segmentation automatic quality control do not predict segmentation quality at slice-level (2D), resulting in sub-optimal evaluations. Our 2D-based deep learning method simultaneously performs quality control at 2D-level and 3D-level for cardiovascular MR image segmentations. We compared it with 3D approaches by training both on 36,540 (2D) / 3842 (3D) samples to predict Dice Similarity Coefficients (DSC) for 4 different structures from the left ventricle, i.e., trabeculations (LVT), myocardium (LVM), papillary muscles (LVPM) and blood (LVC). The 2D-based method outperformed the 3D method. At the 2D-level, the mean absolute errors (MAEs) of the DSC predictions for 3823 samples, were 0.02, 0.02, 0.05 and 0.02 for LVM, LVC, LVT and LVPM, respectively. At the 3D-level, for 402 samples, the corresponding MAEs were 0.02, 0.01, 0.02 and 0.04. The method was validated in a clinical practice evaluation against semi-qualitative scores provided by expert cardiologists for 1016 subjects of the UK BioBank. Finally, we provided evidence that a multi-level QC could be used to enhance clinical measurements derived from image segmentations.