[Paper Review] Fully-automated deep learning slice-based muscle estimation from CT images for sarcopenia assessment

Fahdi Kanavati, Shah Islam+3 more|arXiv (Cornell University)|Jun 10, 2020

Nutrition and Health in Aging4 references4 citations

TL;DR

This study presents a fully automated deep learning framework using fully convolutional neural networks (FCNNs) with a U-Net-like architecture to detect the L3 vertebral slice and segment muscle groups (erector spinae, psoas, rectus abdominus) from abdominal CT scans for sarcopenia assessment. The method achieves high accuracy, with a median slice detection error of 0.50 slices and Dice scores above 0.94 for muscle segmentation, demonstrating performance on par with human experts.

ABSTRACT

Objective: To demonstrate the effectiveness of using a deep learning-based approach for a fully automated slice-based measurement of muscle mass for assessing sarcopenia on CT scans of the abdomen without any case exclusion criteria. Materials and Methods: This retrospective study was conducted using a collection of public and privately available CT images (n = 1070). The method consisted of two stages: slice detection from a CT volume and single-slice CT segmentation. Both stages used Fully Convolutional Neural Networks (FCNN) and were based on a UNet-like architecture. Input data consisted of CT volumes with a variety of fields of view. The output consisted of a segmented muscle mass on a CT slice at the level of L3 vertebra. The muscle mass is segmented into erector spinae, psoas, and rectus abdominus muscle groups. The output was tested against manual ground-truth segmentation by an expert annotator. Results: 3-fold cross validation was used to evaluate the proposed method. The slice detection cross validation error was 1.41+-5.02 (in slices). The segmentation cross validation Dice overlaps were 0.97+-0.02, 0.95+-0.04, 0.94+-0.04 for erector spinae, psoas, and rectus abdominus, respectively, and 0.96+-0.02 for the combined muscle mass. Conclusion: A deep learning approach to detect CT slices and segment muscle mass to perform slice-based analysis of sarcopenia is an effective and promising approach. The use of FCNN to accurately and efficiently detect a slice in CT volumes with a variety of fields of view, occlusions, and slice thicknesses was demonstrated.

Motivation & Objective

• To develop a fully automated, deep learning-based method for slice-based sarcopenia assessment from abdominal CT scans without case exclusions.
• To overcome the time-consuming and inconsistent manual process of L3 slice identification and muscle segmentation in clinical and research workflows.
• To evaluate the performance of the method against expert radiologist annotations using cross-validation and Bland-Altman analysis.
• To enable scalable, reproducible sarcopenia measurement for oncology research and potential clinical screening tools.

Proposed method

• The method uses two separate FCNNs based on a U-Net-like architecture: one for L3 slice detection and one for muscle segmentation.
• Input data consists of 3D CT volumes preprocessed into 2D frontal and sagittal maximal intensity projection (MIP) images to enhance anatomical landmarks.
• Slice detection is trained using confidence maps to localize the L3 vertebra, with restricted sagittal MIPs to improve visibility of the sacrum and spinal alignment.
• Muscle segmentation is performed on the detected L3 slice, with models trained to distinguish erector spinae, psoas, rectus abdominus, and combined muscle mass.
• Both models are trained and evaluated using 3-fold cross-validation on a diverse dataset of 1070 CT volumes from multiple sources.
• The framework is designed to be generalizable to other anatomical levels and adaptable for multi-slice sarcopenia assessment.

Experimental results

Research questions

• RQ1 Can a fully convolutional neural network detect the L3 vertebral slice in abdominal CT volumes with performance comparable to human radiologists?
• RQ2 How accurately can a deep learning model segment individual muscle groups (erector spinae, psoas, rectus abdominus) and combined muscle mass on the L3 slice?
• RQ3 Does the automated sarcopenia measurement derived from the model show statistically significant agreement with manual annotations by expert radiologists?
• RQ4 How does the method perform on challenging cases such as transitional vertebrae or patients with extensive soft-tissue edema?
• RQ5 Can the method be applied to diverse CT scans with varying fields of view, slice thicknesses, and anatomical variations without exclusion criteria?

Key findings

• The slice detection model achieved a median error of 0.50 slices (1.41±5.02 in 3-fold cross-validation), comparable to human inter-radiologist error (median 0.80).
• Muscle segmentation achieved high Dice overlap scores: 0.97±0.02 for erector spinae, 0.95±0.04 for psoas, 0.94±0.04 for rectus abdominus, and 0.96±0.02 for combined muscle mass.
• Bland-Altman analysis showed no statistically significant difference (p=0.9503) between muscle area measurements from the model and a second expert radiologist.
• The method demonstrated robustness on challenging cases, including those with soft-tissue edema and spinal curvature, though some segmentation failures were observed in extreme cases.
• Outlier cases were primarily due to transitional vertebrae or poor spinal visualization, with confidence maps often indicating multiple potential L3 candidates.
• The model’s performance was consistent across diverse CT scans, including those with metal implants, and the framework is generalizable to other anatomical levels.