Get started
Work

Machine Learning Algorithm for Measuring Muscle Mass and Myosteotosis Using Computed Tomography

Year: 2018
Type: article
Abstract: The purpose was to validate a machine learning (ML) algorithm for measuring muscle mass and myosteotosis on computed tomography (CT) examinations. 1404 non-contrast chest CT exams from older participants in the National Lung Screening Trial (mean age 72, mean BMI 27.2, 36% female) were used to validate the algorithm. This dataset was chosen because the CT images are heterogeneous in field of view, voxel spacing, convolution kernel, scanner manufacturer and model, reconstruction algorithm, and image quality. Ground truth was obtained by manually segmenting the left paraspinous muscle at the T12 level with muscle thresholds set at -29 to +150 Hounsfield units (HU). Our ML algorithm for muscle segmentation was developed using open-source tools: Keras, TensorFlow, Pandas, NumPy, scikit-image, Snakemake, and the Oxford Centre for Functional Magnetic Resonance Imaging of the Brain Software Library. The algorithm is a Keras-based implementation of the U-Net convolutional neural network (CNN) architecture, with input images pre-processed into three channels: raw voxel values, clamped and rescaled, and equalized using the CLAHE algorithm. The model was trained over 100 epochs with a batch size of 30, using 690 training images augmented with rotation, translation, and scaling. Pearson correlation and Dice similarity coefficients between ML and manual segmentation were determined. Compared to manual segmentation, the ML algorithm had a median Dice similarity coefficient of 0.91 (SD=0.06). There was high correlation in muscle cross-sectional area (r=0.89; p<0.0001) and muscle attenuation (r=0.92; p<0.0001) between ML algorithm and ground truth.
DOI: 10.1093/geroni/igy031.3302
Language: English
Open Access status: green
OpenAlex ID: W2906830908
FWCI:
Cites:
Cited by:
Related to: