VinDr-Mammo

VinDr-Mammo is an artificial intelligence-powered software tool designed to assist MQSA qualified physicians by prioritizing 2D FFDM screening mammograms that may contain suspicious findings. It analyzes mammography images to flag potentially concerning exams at an exam-level for worklist prioritization within existing PACS/Workstation environments, and its efficacy is supported by validation studies demonstrating high sensitivity and specificity across diverse datasets and imaging systems.

VinDr-Mammo is a software tool designed as a passive notification system for prioritizing 2D full-field digital mammography (FFDM) screening exams. It utilizes an artificial intelligence algorithm to analyze these mammograms and identify exams suggestive of suspicious findings at an exam-level. The software then flags these suspicious cases within a Picture Archiving and Communication System (PACS) or workstation, allowing MQSA qualified interpreting physicians to reorder their worklist and review potentially positive cases sooner. It functions in a parallel workflow, meaning it does not remove any exams from the physician's worklist nor does it provide diagnostic information beyond triage and prioritization; the physician remains responsible for full evaluation and diagnosis according to standard care. The system comprises modules for data input/validation, the AI algorithm for analysis, an API for result transmission, and error handling. The performance of VinDr-Mammo was validated using two main datasets: one from the Radiological Society of North America (RSNA) Screening Mammography Breast Cancer Detection AI Challenge, and another from two Vietnamese hospitals. The RSNA dataset, representing US demographics, showed a sensitivity of 0.889 and specificity of 0.906, with an AUC of 0.958. The Vietnamese dataset, used to demonstrate generalizability across different screening modalities (Hologic, Siemens, Fujifilm, GE), yielded a sensitivity of 0.906 and specificity of 0.911, with an AUC of 0.965. Aggregate results from both datasets showed a sensitivity of 0.900, specificity of 0.910, and an AUC of 0.962. Performance was also shown to be consistent across various subgroups including breast density, age group, scanner model, and lesion type, with an average processing time of 2.8 minutes.

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 VinDr-Mammo algorithm's performance specifically vary across different types of suspicious findings (e.g., masses vs. microcalcifications) beyond the aggregated AUC values presented?; What are the detailed characteristics of the 'independent datasets from various global sites' used for the initial training of the VinDr-Mammo algorithm, and how do they contribute to its generalizability?; Beyond flagging suspicious cases, does the system provide any information on the nature or location of the suspected finding to the interpreting physician within the PACS/Workstation interface?; What is the specific process and criteria for validating new FFDM systems for use with VinDr-Mammo, beyond those listed (GE, Siemens, Fujifilm, Hologic)?; What were the specific 'confounding factors' considered during the construction of the Vietnamese hospital dataset, and how were they mitigated to ensure robust stand-alone detection and triage accuracy?