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

Abstract

Objectives: Body composition assessment using CT images at the L3-level is increasingly applied in cancer research and has been shown to be strongly associated with long-term survival. Robust high-throughput automated segmentation is key to assess large patient cohorts and to support implementation of body composition analysis into routine clinical practice. We trained and externally validated a deep learning neural network (DLNN) to automatically segment L3-CT images.

Methods: Expert-drawn segmentations of visceral and subcutaneous adipose tissue (VAT/SAT) and skeletal muscle (SM) of L3-CT-images of 3187 patients undergoing abdominal surgery were used to train a DLNN. The external validation cohort was comprised of 2535 patients with abdominal cancer. DLNN performance was evaluated with (geometric) dice similarity (DS) and Lin's concordance correlation coefficient.

Results: There was a strong concordance between automatic and manual segmentations with median DS for SM, VAT, and SAT of 0.97 (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: SM 0.964 (0.959-0.968), VAT 0.998 (0.998-0.998), and SAT 0.992 (0.991-0.993). Bland-Altman metrics indicated only small and clinically insignificant systematic offsets; SM radiodensity: 0.23 Hounsfield units (0.5%), SM: 1.26 cm2.m-2 (2.8%), VAT: -1.02 cm2.m-2 (1.7%), and SAT: 3.24 cm2.m-2 (4.6%).

Conclusion: A robustly-performing and independently externally validated DLNN for automated body composition analysis was developed.

Advances in knowledge: This DLNN was successfully trained and externally validated on several large patient cohorts. The trained algorithm could facilitate large-scale population studies and implementation of body composition analysis into clinical practice.

Keywords: CT; body composition; convolutional neural networks; deep learning; image segmentation.

Figure 1.

Segmentation of skeletal muscle (red), visceral adipose tissue (yellow), and subcutaneous adipose tissue (blue) on a single CT slice at the level of the third lumbar vertebra.

Figure 2.

Distribution of geometric dice similarity (DS) on L3 slice for skeletal muscle (SM), subcutaneous fat (SAT), and visceral fat (VAT). (A) Box-whisker plot showing the median DS as the solid horizontal line and the interquartile range as the upper and lower limits of the box. The vertical line ends indicate 1%-tile and 99%-tile, and outliers outside this range are shown as individual dots. (B)-(D) show the distribution of DS as a function of SM area, VAT area, and SAT area, respectively.

Figure 3.

Lin’s concordance correlation (CCC) plots. (A) Skeletal muscle attenuation (SMRA), (B) skeletal muscle index (SMI), (C) visceral fat index (VATI), and (D) subcutaneous fat index (SATI). The units of SMRA are Hounsfield unit (HU). The units of SMI, VATI, and SATI are all cm².m⁻². Reference values were defined as those extracted from human-drawn segmentations. Predicted values were extracted from DLNN-made segmentations. DLNN = deep learning neural network.