A new artificial intelligence (AI) algorithm developed for automated three-dimensional segmentation of adipose tissue and paravertebral muscle on chest computed tomography (CT) scans has shown remarkable feasibility and high accuracy, which could revolutionize body composition analysis.
This retrospective study, part of the Boston Lung Cancer Study conducted from 2000 to 2011, involved 77 patients for adipose tissue quantification and 245 for muscle quantification. The AI algorithm achieved Dice scores exceeding 0.87, demonstrating excellent correlation with traditional manual segmentation. Specifically, for visceral (VAT) and subcutaneous adipose tissue (SAT), the correlation coefficients surpassed 0.98, indicating the algorithm's reliability.
Computed tomography is predominantly used for diagnosing conditions related to lung and mediastinal illness. Yet, it also offers extensive data about other structures, including adipose tissue and muscle mass. Body composition analysis (BCA), which quantifies these elements, is pivotal for evaluating nutritional status, managing diseases like obesity and sarcopenia, and predicting outcomes, especially for cancer patients.
Sarcomas are marked by muscle loss, impacting rehabilitation and recovery. Therefore, developing precise methods to assess muscle quality and quantity through imaging technology is indispensable. By employing the AI-driven algorithm, researchers can now automate the otherwise labor-intensive process of 3D segmentation, reducing the need for human intervention.
Patients with newly diagnosed stage I non-small cell lung cancer (NSCLC) formed the study cohort. Chest CT scans from these patients were evaluated to differentiate between VAT and SAT accurately. The study established rigorous criteria for assessing image quality and included only non-enhanced CT scans for adipose tissue measurement to minimize variable influences from imaging parameters.
The study also utilized convolutional neural networks (CNN), particularly the U-Net architecture, to build the segmentation model. Through structured training, the AI became proficient at isol distinguishing adipose regions based on established Hounsfield Unit thresholds, allowing for precise volumetric quantification.
Statistical analysis of the segmented results revealed mean differences between the AI and manual segmentation were minimal. For VAT/SAT ratios, the mean absolute difference was 0.7%. Researchers noted the established algorithms could manage variability from different slice thicknesses, improving efficiency and accuracy compared to manual segmentation methods.
Additional insights include the small sample size limitation, which could affect the generalizability of the findings. Despite this, correlations between adipose measurements and clinical outcomes have been established, emphasizing the utility of body composition metrics as prognostic indicators.
With potential applications across diverse medical fields, including oncology and nutrition, the feasibility of applying this AI algorithm could lead to significant advancements. It opens pathways for more consistent, reliable assessments of body composition, enhancing clinical workflows and patient care strategies.
Moving forward, continued exploration of this technology could augment its applicability for large-scale usage and perhaps even real-time imaging assessments during clinical encounters. The study’s researchers advocate for expanded datasets and improved testing parameters to develop more refined and comprehensive algorithms.