||1R01CA277739-01 Interpret this number
||University Of Houston
||SCH: Ai-Doctor Collaborative Medical Diagnosis
Recent retrospective studies show that radiology's diagnostic error rates did not decrease significantly over
the years. For example, missed lung cancer rates remain at 20-60% on chest radiography dependent on
study design. This error contributes to 40,000-80,000 deaths annually in U.S. hospitals. This project aims
to develop a computational framework for Al to collaborate with human radiologists on medical diagnosis
tasks. To achieve this goal, we divide the project into three Aims, where the first two focus on fundamental
theories, and the last one evaluates the proposed approaches on targeted applications.
Aim 1: Develop computational principles for optimal Al-radiologist interaction. This Aim will develop
a computational framework for guiding the interaction between radiologists and Al to achieve the best
possible diagnostic performance while minimizing the time burden. Our framework consists of the first
method for reverse-engineering radiologists' intention from the joint gaze and visual information based on
reinforcement learning. This Aim is the first to provide an integrated system with gaze sensing, deep
networks, and human radiologists. The knowledge from this Aim will fundamentally transform how one
would build collaborative medical diagnosis systems.
Aim 2: Design a user-friendly and minimally-interfering interface for radiologist-Al interaction. This
Aim addresses an essential question of designing a minimally interfering interface that allows human
radiologists to interact with Al models efficiently. Our proposed system combines an innovative "multimodal
thinking with audio and gaze" (MTAG) methodology with user-centered iterative design. The process will
result in a novel radiologist-Al collaborative interface that maximizes time efficiency while minimizing the
amount of distraction. The outcome of this Aim will shed light on design principles for systems involving
Aim 3: Evaluation Plan. This Aim evaluates the proposed approaches in Aim 1-2 on two clinically
important applications: i) Lung nodule detection and ii) pulmonary embolism. Lung cancer is the second
most common cancer, and pulmonary embolism is the third most common cause of cardiovascular death.
Studying how radiologists collaborate with Al to reduce diagnostic errors will lead to significant clinical
RELEVANCE (See instructions):
Diagnostic errors contribute to 40,000-80,000 deaths annually in U.S. hospitals. This project combines
novel artificial intelligence (Al) algorithms, gaze monitoring software, and design principles to help doctors
minimize diagnostic errors due to cognitive and perceptual biases. The project's success will fundamentally
change how we design Al medical systems to increase diagnostic accuracy, save lives, reduce missed
cancer diagnoses, improve public health, and advance NCl's mission.