Abdominal muscle segmentation from CT using a convolutional neural network

A deep learning model based on the attention mechanism for automatic segmentation of abdominal muscle and fat for body composition assessment

Background

Quantitative muscle and fat data obtained through body composition analysis are expected to be a new stable biomarker for the early and accurate prediction of treatment-related toxicity, treatment response, and prognosis in patients with lung cancer. The use of these biomarkers can enable the adjustment of individualized treatment regimens in a timely manner, which is critical to further improving patient prognosis and quality of life. We aimed to develop a deep learning model based on attention for fully automated segmentation of the abdomen from computed tomography (CT) to quantify body composition.

Methods

A fully automatic segmentation deep learning model was designed based on the attention mechanism and using U-Net as the framework. Subcutaneous fat, skeletal muscle, and visceral fat were manually segmented by two experts to serve as ground truth labels. The performance of the model was evaluated using Dice similarity coefficients (DSCs) and Hausdorff distance at 95th percentile (HD95).

Results

The mean DSC for subcutaneous fat and skeletal muscle were high for both the enhanced CT test set (0.93±0.06 and 0.96±0.02, respectively) and the plain CT test set (0.90±0.09 and 0.95±0.01, respectively). Nevertheless, the model did not perform well in the segmentation performance of visceral fat, especially for the enhanced CT test set. The mean DSC for the enhanced CT test set was 0.87±0.11, while the mean DSC for the plain CT test set was 0.92±0.03. We discuss the reasons for this result.

Conclusions

This work demonstrates a method for the automatic outlining of subcutaneous fat, skeletal muscle, and visceral fat areas at L3.

Transfer learning for data-efficient abdominal muscle segmentation with convolutional neural networks

Background

Skeletal muscle segmentation is an important procedure for assessing sarcopenia, an emerging imaging biomarker of patient frailty. Data annotation remains the bottleneck for training deep learning auto‐segmentation models.

Purpose

There is a need to define methodologies for applying models to different domains (e.g., anatomical regions or imaging modalities) without dramatically increasing data annotation.

Methods

To address this problem, we empirically evaluate the generalizability of various source tasks for transfer learning: natural image classification, natural image segmentation, unsupervised image reconstruction, and self‐supervised jigsaw solving. Axial CT slices at L3 were extracted from PET‐CT scans for 204 oesophago‐gastric cancer patients and the skeletal muscle manually delineated by an expert. Features were transferred and segmentation models trained on subsets (n=5,10,25,50,75,100,125) of the manually annotated training set. Four‐fold cross‐validation was performed to evaluate model generalizability. Human‐level performance was established by performing an inter‐observer study consisting of ten trained radiographers.

Results

We find that accurate segmentation models can be trained on a fraction of the data required by current approaches. The Dice similarity coefficient and root mean square distance‐to‐agreement were calculated for each prediction and used to assess model performance. Models pre‐trained on a segmentation task and fine‐tuned on 10 images produce delineations that are comparable to those from trained observers and extract reliable measures of muscle health.

Conclusions

Appropriate transfer learning can generate convolutional neural networks for abdominal muscle segmentation that achieve human‐level performance while decreasing the required data by an order of magnitude, compared to previous methods (n=160→10). This work enables the development of future models for assessing skeletal muscle at other anatomical sites where large annotated data sets are scarce and clinical needs are yet to be addressed.

Validation for measurements of skeletal muscle areas using low-dose chest computed tomography

Various methods were suggested to measure skeletal muscle areas (SMAs) using chest low-dose computed tomography (chest LDCT) as a substitute for SMA at 3rd lumbar vertebra level (L3-SMA). In this study, four SMAs (L1-SMA, T12-erector spinae muscle areas, chest wall muscle area at carina level, pectoralis muscle area at aortic arch level) were segmented semi-automatically in 780 individuals taking concurrent chest and abdomen LDCT for healthcare screening. Four SMAs were compared to L3-SMA and annual changes were calculated from individuals with multiple examinations (n = 101). Skeletal muscle index (SMI; SMA/height²) cut-off for sarcopenia was determined by lower 5th percentile of young individuals (age ≤ 40 years). L1-SMA showed the greatest correlation to L3-SMA (men, R² = 0.7920; women, R² = 0.7396), and the smallest annual changes (0.3300 ± 4.7365%) among four SMAs. L1-SMI cut-offs for determining sarcopenia were 39.2cm²/m² in men, and 27.5cm²/m² in women. Forty-six men (9.5%) and ten women (3.4%) were found to have sarcopenia using L1-SMI cut-offs. In conclusion, L1-SMA could be a reasonable substitute for L3-SMA in chest LDCT. Suggested L1-SMI cut-offs for sarcopenia were 39.2cm²/m² for men and 27.5cm²/m² for women in Asian.

Deep learning-based quantification of temporalis muscle has prognostic value in patients with glioblastoma

Background

Glioblastoma is the commonest malignant brain tumour. Sarcopenia is associated with worse cancer survival, but manually quantifying muscle on imaging is time-consuming. We present a deep learning-based system for quantification of temporalis muscle, a surrogate for skeletal muscle mass, and assess its prognostic value in glioblastoma.

Methods

A neural network for temporalis segmentation was trained with 366 MRI head images from 132 patients from 4 different glioblastoma data sets and used to quantify muscle cross-sectional area (CSA). Association between temporalis CSA and survival was determined in 96 glioblastoma patients from internal and external data sets.

Results

The model achieved high segmentation accuracy (Dice coefficient 0.893). Median age was 55 and 58 years and 75.6 and 64.7% were males in the in-house and TCGA-GBM data sets, respectively. CSA was an independently significant predictor for survival in both the in-house and TCGA-GBM data sets (HR 0.464, 95% CI 0.218–0.988, p = 0.046; HR 0.466, 95% CI 0.235–0.925, p = 0.029, respectively).

Conclusions

Temporalis CSA is a prognostic marker in patients with glioblastoma, rapidly and accurately assessable with deep learning. We are the first to show that a head/neck muscle-derived sarcopenia metric generated using deep learning is associated with oncological outcomes and one of the first to show deep learning-based muscle quantification has prognostic value in cancer.

Deep-Learning Assessed Muscular Hypodensity Independently Predicts Mortality in DLBCL Patients Younger Than 60 Years

Simple Summary

Cachexia is a major cause of mortality in cancer patients and is characterized by a continuous skeletal muscle loss. Muscle depletion assessed by computed tomography (CT) is a predictive marker in solid tumors but has never been assessed in non-Hodgkin’s lymphoma. Despite software improvements, its measurement remains highly time-consuming and cannot be performed in clinical practice. We report the development of a CT segmentation algorithm based on convolutional neural networks. It automates the extraction of anthropometric data from pretherapeutic CT to assess precise body composition of young diffuse large B cell lymphoma (DLBCL) patients at the time of diagnosis. In this population, muscle hypodensity appears to be an independent risk factor for mortality, and can be estimated at diagnosis with this new tool.

Abstract

Background. Muscle depletion (MD) assessed by computed tomography (CT) has been shown to be a predictive marker in solid tumors, but has not been assessed in non-Hodgkin’s lymphomas. Despite software improvements, MD measurement remains highly time-consuming and cannot be used in clinical practice. Methods. This study reports the development of a Deep-Learning automatic segmentation algorithm (DLASA) to measure MD, and investigate its predictive value in a cohort of 656 diffuse large B cell lymphoma (DLBCL) patients included in the GAINED phase III prospective trial (NCT01659099). Results. After training on a series of 190 patients, the DLASA achieved a Dice coefficient of 0.97 ± 0.03. In the cohort, the median skeletal muscle index was 50.2 cm²/m² and median muscle attenuation (MA) was 36.1 Hounsfield units (HU). No impact of sarcopenia was found on either progression free survival (PFS) or overall survival (OS). Muscular hypodensity, defined as MA below the tenth percentile according to sex, was associated with a lower OS and PFS, respectively (HR = 2.80 (95% CI 1.58–4.95), p < 0.001, and HR = 2.22 (95% CI 1.43–3.45), p < 0.001). Muscular hypodensity appears to be an independent risk factor for mortality in DLBCL and because of DLASA can be estimated in routine practice.