ClariPulmo

ClariPulmo is an image analysis software that utilizes deep learning for automated lung segmentation and quantifies low and high-attenuation areas in thoracic CT images using Hounsfield unit thresholds to support physician diagnosis. It enhances CT image quantification through optional U-Net based denoising and kernel normalization functions, particularly for low-dose or sharp kernel images.

ClariPulmo is a non-invasive, standalone software application designed to support physicians in the quantification of lung CT images for examining pulmonary tissues and abnormalities. Its core functionality involves automated segmentation of the lungs using a pre-trained deep learning model. Following segmentation, the software quantifies low-attenuation areas (LAA) and high-attenuation areas (HAA) within the lungs by counting voxels based on user-predefined Hounsfield unit (HU) thresholds. The analysis results, including color overlay images of the segmented lungs, tables, and charts, are provided in summary reports to aid in diagnosis and documentation, although it is explicitly not intended for diagnosing pneumonia or COVID-19. The software incorporates optional advanced features to improve image quality and quantification accuracy, particularly for challenging CT datasets. These include an AI-based kernel normalization function, built on U-Net architecture, which translates images from sharp to smooth kernels, and an AI-based denoising function, also using U-Net architecture, to reduce noise in low-dose CT (LDCT) images. Performance testing demonstrated excellent agreement of its HAA analysis (PCC: 0.980–0.983) and LAA analysis (PCC: 0.99) with expert-established segmentations. The AI-based lung segmentation also showed high accuracy (PCC: 0.977–0.992; DICE coefficients: 0.98–0.99) across various patient datasets, CT scanners, reconstruction kernels, and low-dose conditions when compared to expert radiologist segmentations.

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. What specific features or architectural modifications of the U-Net are employed for the denoising and kernel normalization tasks, and how were these models trained and validated?; How does the performance of the automated lung segmentation algorithm vary across different types of interstitial lung diseases or in the presence of extensive pleural abnormalities?; What is the range of user-predefined Hounsfield unit thresholds typically used for LAA and HAA analysis, and how does the software guide users in selecting appropriate thresholds?; Can the ClariPulmo software integrate with existing PACS systems, and what is the typical workflow for a radiologist using this tool in a clinical setting?; Beyond PCC and DICE coefficients, what other validation metrics were used to assess the accuracy and robustness of the LAA, HAA, and lung segmentation features, particularly concerning inter-scanner variability?