CT Cardiomegaly

CT Cardiomegaly is a command-line software application that employs a non-adaptive machine learning algorithm to automatically analyze axial CT slices by segmenting the heart and thoracic cavity. Its primary function is to calculate both linear and area-based cardiothoracic ratios, providing quantitative data to assist clinicians in assessing cardiac size, with the explicit understanding that its results are not intended as a sole basis for diagnosis.

CT Cardiomegaly is a software-only medical device, functioning as a command-line application, designed to automatically calculate both linear and area-based cardiothoracic ratios (CTR) from computed tomography (CT) images that include the heart. It utilizes a non-adaptive machine learning algorithm, built using the MONAI framework, to identify and segment the heart and inner chest cavity from an axial CT slice. The software measures the maximal transverse diameter of the heart and the maximal inner transverse diameter of the thoracic cavity to derive these CTR values. Intended for use by physicians and licensed practitioners in various healthcare settings, it processes CT studies from adult and transitional adolescent patients (18-21 years old) and outputs results in PDF and JSON formats. These reports include the calculated ratios, patient information, and user-configurable reference values, but are not intended for standalone clinical decision-making. The machine learning algorithm within CT Cardiomegaly was trained using a reference standard derived from segmentations by three board-certified radiologists, with a training dataset of 1600 unique patient cases from 267 sites, ensuring no site overlap with the testing cohort which comprised 275 unique patient cases. This testing dataset was diverse, including varied patient demographics, confounding conditions, and imaging characteristics from numerous scanner manufacturers and protocols. Performance assessments demonstrated strong agreement with physician-derived measurements, achieving a mean difference of less than 0.01 for both Cardiothoracic Ratio (linear) and Heart to Chest Area Ratio (area-based). The 95% limits of agreement were reported as [-0.033, 0.035] for linear CTR and [-0.028, 0.035] for area-based CTR. Secondary endpoints, such as key heart slice detection accuracy (mean difference 0.46mm) and segmentation quality (Dice scores of 0.95 for heart and 0.98 for inner chest), further validated the device's analytical performance.

AI Regulatory Concerns

Powered by

Jev reads the full FDA letter and returns calibrated concern probabilities per regulatory category. These are AI estimates, not regulatory advice.

• Q1. How does the non-adaptive nature of the machine learning algorithm affect its performance on patient populations or imaging protocols significantly different from the training dataset of 1600 cases?; What specific image processing steps or algorithmic features are employed to reliably identify the optimal axial CT slice for CTR calculation from a full CT volume, especially in cases with anatomical variations or suboptimal image quality?; Given the device outputs both linear and area-based CTR, what is the comparative clinical sensitivity and specificity of area-based CTR versus linear-based CTR for detecting cardiomegaly when derived from CT images, and in what specific clinical scenarios might one be preferred over the other?; The performance data indicates strong agreement with human readers; however, what are the detailed characteristics of the 275 test cases, particularly regarding the prevalence and types of cardiomegaly or other co-existing thoracic pathologies, and how did the device perform across these specific subgroups?; How does the MONAI framework contribute to the algorithm's architecture and performance, and are there plans to update the non-adaptive model as new data or improved machine learning techniques become available?