Transpara 1.7.2

Transpara 1.7.2 is a software medical device utilizing deep learning algorithms to assist physicians in the detection and diagnosis of breast cancer from FFDM and DBT images. It aims to improve interpretive accuracy by providing computer-aided detection marks and quantitative scores for suspicious regions and overall exams, based on algorithms trained and validated on diverse radiological imaging data.

Transpara 1.7.2 is a radiological computer-assisted detection and diagnosis software designed to aid physicians in interpreting screening full-field digital mammography (FFDM) and digital breast tomosynthesis (DBT) exams. Its primary purpose is to identify regions suspicious for breast cancer and assess their likelihood of malignancy, serving as a concurrent reading aid. The system processes DICOM images from compatible FFDM and DBT systems, applying 'deep learning' algorithms trained on an extensive database of biopsy-proven breast cancers, benign abnormalities, and normal tissue. The software outputs computer-aided detection (CAD) marks for suspicious calcifications and soft-tissue lesions, region scores (0-100 scale indicating suspicion level), links between corresponding regions in different views, and an overall exam score (1-10 scale for cancer likelihood in the exam). Non-clinical standalone performance tests were conducted using an independent dataset of 10,690 exams (5,867 FFDM, 4,823 DBT), which included 1,472 cancer cases with varied morphologies (e.g., 56.5% mass, 26.1% calcifications) and histological types. For FFDM, Transpara 1.7.2 demonstrated a sensitivity of 95.0% at 0.30 false positives per image, with an Area Under the Curve (AUC) of 0.945. For DBT, it achieved a sensitivity of 93.2% at 0.34 false positives per volume, with an AUC of 0.945. These tests showed its detection performance to be non-inferior to its predicate (Transpara 1.7.0); it is intended for use by qualified physicians as an adjunct tool, and patient management decisions should not be based solely on its analysis.

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 'deep learning' algorithm in Transpara 1.7.2 differentiate between various histological subtypes of breast cancer, and does its performance vary significantly among them?; What specific architectural changes were implemented in the algorithms for version 1.7.2 compared to 1.7.0 to achieve improved detection accuracy for DBT, particularly concerning different lesion types?; How does the exam score calibration, which aims for approximately 10 percent of non-cancer mammograms to fall into each of the 10 risk categories, account for variability in breast density and patient demographics across diverse populations?; What is the specific performance (sensitivity, false positive rate) of Transpara 1.7.2 in identifying architectural distortions and asymmetries, considering they represented only 4.0% of cancer regions in the test dataset?; Given that the user interface implementation is by third parties, what are the key technical requirements or best practices for integrating Transpara's DICOM output effectively into various PACS and medical workstations to optimize radiologist workflow?