Toward efficient performance for deep learning on medical imaging

University Of California, San Francisco, San Francisco CA

Investigators

Linked publications, trials & patents

Abstract

Project Summary / Abstract Objective — The goal of this renewal is to continue developing and optimizing novel deep learning (DL) and related methods to improve diagnosis and clinical decision-making for congenital heart disease (CHD). Nota- bly, this work includes a clinical translational evaluation of these methods in a population-wide imaging collec- tion spanning tens of thousands of patients and several clinical centers. Background — Despite clear and numerous benefits to prenatal detection of CHD and an ability for fetal ultrasound to detect over 90% of CHD in theory, in practice detection is closer to 50%. Literature suggests a key cause of this startling diagnosis gap is suboptimal image acquisition and interpretation. Progress — In progress to date, our multi-disciplinary team in ultrasound, CHD and machine learning have pioneered the use of DL for improving CHD detection from sec- ond-trimester fetal screening ultrasound. In a landmark paper, we achieved an AUC of 0.99 on over 4,500 screening ultrasounds, setting a new standard for evaluating DL by testing on over 400 times more images than the model was trained on. In this renewal, we will study advances in DL to push the limits of generalizability and robustness yet again, to build a solution robust enough to scale to the world's hundreds of millions of fetal ultrasounds annually. Preliminary Work — We have developed DL models to perform cross- functional tasks on fetal imaging (including classification, segmentation, and object detection) that together can inform more efficient learning. We have leveraged semi- and self-supervised learning techniques to reduce reli- ance on laborious and error-prone manual labeling. We have used clinical knowledge in a DL pipeline to pro- duce more robust results (although not yet directly integrated into the models themselves). We have proto- typed software that assists clinicians in labeling images and evaluating DL predictions. Hypothesis — Accu- racy and robustness for challenging, truly large-scale (>200M/year) ultrasound tasks such as CHD detection will require stronger learned features achieved by incorporating clinical knowledge directly into DL model train- ing. Robust evaluation will require tools to allow clinicians across medical centers to participate. Aims — To this end, we propose (1) to develop and optimize more robust feature extraction and stronger learned features for CHD detection, (2) to develop and optimize neural networks that can directly incorporate clinical knowledge (“priors”) into training, and (3) to develop for the research community open-source, customizable software for secure, cloud-based clinician evaluation of imaging tasks. We will benchmark improvement over work to date by demonstrating expert-level performance on large, external community imaging datasets, on first-trimester imaging; and by developing novel biometrics. Environment and Impact — This work proposed takes place in an outstanding environment at the crossroads of data science, cardiovascular and fetal imaging, and transla- tional informatics. It will provide valuable tools and insight into designing and evaluating DL models on imaging for clinically relevant goals at clinical-grade performance and scale.