DeepHealth ProstateAI

DeepHealth ProstateAI is a computer-aided detection and diagnosis (CADe/x) software designed to assist trained physicians in detecting and characterizing prostate cancer lesions on MRI scans in patients aged 40 and older. It analyzes T2-weighted, ADC, and DWI MR images combined with a binary prostate segmentation mask to identify regions of interest and classify them as low, moderate, or high suspicion for cancer. The device is intended as a concurrent reading aid alongside physician interpretation, not as a replacement for clinical decision-making.

DeepHealth ProstateAI uses a convolutional neural network (CNN) based on 3D Retina U-Net within the nnDetection framework, whereas the predicate ProstatID uses Random Forest models. Both analyze T2W, DWI, and ADC sequences to detect and characterize suspicious lesions. DeepHealth outputs ROI-level segmentations with categorical classifications (low, moderate, high) as DICOM Segmentation Objects or NIfTI files, while ProstatID outputs continuous malignancy likelihood scores and PI-RADS-suggestive results. Both are validated for Philips, GE, and Siemens MRI systems (1.5T and 3T) and integrate into PACS workflows.

FDA 21 CFR 820 (Quality System Regulation); ISO 14971:2019 (Risk Management); IEC 62304:2015 (Software Life Cycles); IEC 62366:2020 (Usability Engineering); IEC 82304-1:2016 (Health Software Safety); NEMA PS3 (DICOM); FDA guidance on SaMD Clinical Evaluation (December 2017), Medical Device Use Safety (February 2016), and Cybersecurity in Medical Devices (June 2025).

DeepHealth ProstateAI is substantially equivalent to ProstatID because both are radiological CADe/x software devices (21 CFR 892.2090, Product Code QDQ) intended to aid qualified physicians in detecting and characterizing prostate cancer on MRI. Although the underlying algorithms differ (CNN vs. Random Forest) and output formats vary (categorical vs. continuous scores), both serve the same clinical purpose, process identical imaging sequences, and are validated across comparable scanner platforms. Clinical performance testing demonstrated statistically significant improvement in detection (AUC-LROC 0.626 to 0.724, p<0.05) and met predefined acceptance criteria, with no new questions of safety or effectiveness raised.