Validation of a deep learning model for automatic segmentation of skeletal muscle and adipose tissue on L3 abdominal CT images

Background Body composition assessment using abdominal computed tomography (CT) images is increasingly applied in clinical and translational research. Manual segmentation of body compartments on L3 CT images is time-consuming and requires significant expertise. Robust high-throughput automated segmentation is key to assess large patient cohorts and ultimately, to support implementation into routine clinical practice. By training a deep learning neural network (DLNN) with several large trial cohorts and performing external validation on a large independent cohort, we aim to demonstrate the robust performance of our automatic body composition segmentation tool for future use in patients. Methods L3 CT images and expert-drawn segmentations of skeletal muscle, visceral adipose tissue, and subcutaneous adipose tissue of patients undergoing abdominal surgery were pooled (n = 3,187) to train a DLNN. The trained DLNN was then externally validated in a cohort with L3 CT images of patients with abdominal cancer (n = 2,535). Geometric agreement between automatic and manual segmentations was evaluated by computing two-dimensional Dice Similarity (DS). Agreement between manual and automatic annotations were quantitatively evaluated in the test set using Lin’s Concordance Correlation Coefficient (CCC) and Bland-Altman’s Limits of Agreement (LoA). Results The DLNN showed rapid improvement within the first 10,000 training steps and stopped improving after 38,000 steps. There was a strong concordance between automatic and manual segmentations with median DS for skeletal muscle, visceral adipose tissue, and subcutaneous adipose tissue of 0.97 (interquartile range, IQR: 0.95-0.98), 0.98 (IQR: 0.95-0.98), and 0.95 (IQR: 0.92-0.97), respectively. Concordance correlations were excellent: skeletal muscle 0.964 (0.959-0.968), visceral adipose tissue 0.998 (0.998-0.998), and subcutaneous adipose tissue 0.992 (0.991-0.993). Bland-Altman metrics (relative to approximate median values in parentheses) indicated only small and clinically insignificant systematic offsets : 0.23 HU (0.5%), 1.26 cm 2 .m -2 (2.8%), -1.02 cm 2 .m -2 (1.7%), and 3.24 cm 2 .m -2 (4.6%) for skeletal muscle average radiodensity, skeletal muscle index, visceral adipose tissue index, and subcutaneous adipose tissue index, respectively. Assuming the decision thresholds by Martin et al. for sarcopenia and low muscle radiation attenuation, results for sensitivity (0.99 and 0.98 respectively), specificity (0.87 and 0.98 respectively), and overall accuracy (0.93) were all excellent. Conclusion We developed and validated a deep learning model for automated analysis of body composition of patients with cancer. Due to the design of the DLNN, it can be easily implemented in various clinical infrastructures and used by other research groups to assess cancer patient cohorts or develop new models in other fields.