ViewFinder Software Version 1.1

ViewFinder Software Version 1.1 is designed to assist clinicians in reviewing Digital Breast Tomosynthesis (DBT) images by dynamically correlating tissue between different views using a two-stage process involving simulated tissue movement and a locked Artificial Intelligence algorithm. Its methodology relies on proprietary AI/ML trained on diverse clinical cases to improve tissue matching accuracy, aiming to reduce cognitive load and support efficient screening and diagnostic interpretation by qualified physicians.

ViewFinder Software Version 1.1 is a medical image management and processing system designed to enhance the review of Digital Breast Tomosynthesis (DBT) images. Its primary function is to display two DBT views of the same breast, typically Cranio-Caudal (CC) and Medio-Lateral Oblique (MLO) views, and dynamically indicate correlated or matched tissue between them, aiming to help clinicians compare tissue more efficiently and with reduced cognitive effort. The software achieves this by first simulating tissue movements between the different compressions for a gross approximation of tissue matching, followed by a fine-tuning step utilizing a locked Artificial Intelligence (AI) algorithm. ViewFinder operates as a standalone application or can be integrated into existing image management systems, running on a PC with mammography-cleared monitors, and is intended for use by qualified interpreting physicians in both screening and diagnostic DBT review, providing visualization and image enhancement tools while the physician remains responsible for the final diagnosis. The software's tissue registration capabilities are based on proprietary AI and Machine Learning algorithms, which are 'locked,' meaning their predictive outputs do not change in the field and are only updated through new software versions. These algorithms were trained on a large, diverse dataset of tomosynthesis cases with ground-truth annotations provided by a consultant radiologist, with no overlap between training and test data. Non-clinical testing and scientific data support its functionality, with performance evaluated using metrics such as the correct mapping percentage (frequency of ground truth tissue falling within the predicted area) and Total Registration Errors (TRE). This evaluation was conducted on a test set of 34 lesions from 30 diverse cases, selected for a high prevalence of difficult and cancerous conditions, with performance stratified by breast density and feature size, generally exceeding a 70% minimum performance threshold.

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 performance of the locked AI algorithm in ViewFinder for tissue correlation vary across different DBT acquisition systems not included in the initial supported list (Hologic Selenia Dimensions, GE SenoClaire, GE Senographe Pristina)?; What is the quantitative impact of ViewFinder's tissue correlation feature on reading time and diagnostic accuracy for radiologists with varying levels of experience in DBT interpretation?; The document mentions 'a gross approximation of tissue matching' followed by 'fine tuning using locked AI.' Can the specific contributions of each stage to the final registration accuracy be quantified, and what are the failure modes of the gross approximation step?; Given the training data was sourced from assessment clinics with a high prevalence of difficult and cancerous cases, how does the algorithm perform on a more general screening population with a lower prevalence of such findings?; How does the Total Registration Error (TRE) distribution for ViewFinder compare to other existing DBT image registration methods, and what is the clinical significance of the reported TRE values for different lesion types and sizes?