Body composition assessment by artificial intelligence from routine computed tomography scans in colorectal cancer: Introducing BodySegAI

Abstract

Background Body composition is of clinical importance in colorectal cancer patients, but is rarely assessed because of time-consuming manual segmentation. We developed and tested BodySegAI, a deep learning-based software for automated body composition quantification from routinely acquired computed tomography (CT) scans. Methods A two-dimensional U-Net convolutional network was trained on 2989 abdominal CT slices from L2 to S1 to segment skeletal muscle (SM), visceral adipose tissue (VAT), subcutaneous adipose tissue (SAT), and intermuscular and intramuscular adipose tissue (IMAT). Human ground truth was established by combining segmentations from three human readers. BodySegAI was tested using 154 slices against the human ground truth and compared with a software named AutoMATiCA. Results Median Dice scores for BodySegAI against human ground truth were 0.969, 0.814, 0.986, and 0.990 for SM, IMAT, VAT, and SAT, respectively. The mean differences per slice for SM were À0.09 cm 3 , IMAT: À0.17 cm 3 , VAT: À0.12 cm 3 , and SAT: 0.67 cm 3 . Median absolute errors for SM, IMAT, VAT, and SAT were 1.35, 10.54, 0.91, and 1.07%, respectively. When analysing different anatomical levels separately, L3 and S1 demonstrated the overall highest and lowest Dice scores, respectively. On average, BodySegAI segmented 148 times faster than human readers (4.9 vs. 726.5 seconds, P < 0.001). Also, BodySegAI presented higher Dice scores for SM, IMAT, SAT, and VAT than AutoMATiCA (slices = 154). Conclusions BodySegAI rapidly generates excellent segmentation of SM, VAT, and SAT and good segmentation of IMAT in L2 to S1 among colorectal cancer patients and may replace semi-manual segmentation.

A systematic review of automated segmentation of 3D computed‐tomography scans for volumetric body composition analysis

“…Ground‐truth BC labelling was performed by a radiologist, a radiographer and a dietician. The authors then used the STAPLE (simultaneous truth and performance level estimation) algorithm to generate the optimum single ground‐truth by combining segmentations from all three labellers 57 . This algorithm originates from neuroradiology as a means of establishing optimal ground‐truth labels from multiple labellers for training of neural networks to segment brain tumours.…”

“…The authors then used the STAPLE (simultaneous truth and performance level estimation) algorithm to generate the optimum single ground-truth by combining segmentations from all three labellers. 57 This algorithm originates from neuroradiology as a means of establishing optimal ground-truth labels from multiple labellers for training of neural networks to segment brain tumours. Briefly, the STAPLE algorithm applies weights to each labeller's manual segmentations, by assessing their accuracy.…”

Human parasagittal dura is a potential neuroimmune interface

“…Most large veins were excluded in the previous threshold step, but some high-signal vessels had to be removed manually. An in-house artificial intelligence (AI)-software based on a two-dimensional U-Net 43 convolutional network with close resemblance to a tested model, BodySegAI 44 , was trained from this data and used to segment PSD in another seven subjects. U-Net was chosen for good performance from a limited training set.…”