AI-Rad Companion (Pulmonary)

AI-Rad Companion (Pulmonary) SW version VA40 is an advanced image processing software designed to support clinicians in the assessment of lung diseases by providing quantitative and qualitative analysis of CT DICOM images. It utilizes machine learning and deep learning algorithms for functionalities such as lung and lobe segmentation, identification of Hounsfield value alterations, and comprehensive segmentation and follow-up of both solid and sub-solid pulmonary nodules.

AI-Rad Companion (Pulmonary) SW version VA40 is image processing software designed to assist clinicians in evaluating lung diseases using previously acquired Computed Tomography (CT) DICOM images. Employing machine learning and deep learning algorithms, it provides quantitative and qualitative analysis, including segmentation and measurement of lungs, lung lobes, and both solid and sub-solid lung nodules. The software identifies regions with altered Hounsfield values, dedicates nodules to specific lobes, correlates current findings with prior scans to assess changes, and can interface with external CAD systems. It is intended for use with adult patients (22 years and older) and has been validated with data from Siemens Healthineers, GE Healthcare, and Philips CT scanners, functioning as an adjunct tool for qualified medical professionals and not replacing their judgment. A key enhancement in this version (VA40) is the introduction of sub-solid lung nodule segmentation, expanding its lesion analysis capabilities beyond the previous version which primarily handled solid nodules. The software's performance was validated through non-clinical testing against its predicate device, using data from 527 subjects across the United States and Germany. Results demonstrated significantly better performance than the predicate for existing solid and calcified nodule types, as indicated by a higher average DICE coefficient. Furthermore, the new sub-solid nodule segmentation feature showed superior DICE coefficient performance compared to the predicate's solid nodule segmentation. The validation process included rigorous ground truth annotation by experienced radiologists and ensured independence between training and testing datasets, confirming the software's reliability and improved accuracy.

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 algorithm differentiate between various sub-types of sub-solid nodules (e.g., pure ground-glass, part-solid), and what is its performance accuracy for each sub-type?; What specific machine learning/deep learning architectures are employed for the lung lobe and nodule segmentation tasks, and how were they optimized for CT images from different vendors and reconstruction techniques?; The validation data included low-dose and conventional-dose CT scans; are there statistically significant differences in the software's performance between these two dose levels, particularly for small or sub-solid nodules?; Regarding lesion follow-up, what is the minimum detectable change in nodule size or volume that the software can reliably quantify, and how does this compare to inter-observer variability among radiologists?; Given the validation on images from Siemens, GE, and Philips, what is the anticipated performance or potential limitations when processing CT images from other manufacturers or newer reconstruction algorithms not included in the validation set?