Multicentre study, Not applicable, Retrospective, Cancer, Segmentation, CAD, CT, Artificial Intelligence, Abdomen, Artificial Intelligence and Machine Learning
P. Moeskops1, B. de Vos2, W. B. Veldhuis1, P. A. de Jong1, I. Išgum2, T. Leiner1; 1Utrecht/NL, 2Amsterdam/NL
Methods and materials
150 native full body CTs were used (120 kVp, 5 – 200 mAs, in-plane resolution 0.86 – 1.03 mm, resampled to 5 mm slices). Trained observers defined the reference standard by manual annotation of subcutaneous fat, visceral fat, psoas muscle, abdominal muscle and long spine muscles at L3 level. 100 images were used to train two convolutional neural networks: the first network detects the L3 slice from the 3D volume and the second network segments the 5 tissue classes in the detected slice. The remaining 50 images were used to evaluate the performance of the method. Slice selection performance was evaluated as the distance between manual and automatically selected slices. Segmentation performance was evaluated with Dice coefficients between manual and automatic segmentations.