Computer Vision Technologies for Rapid Detection of the Acute Respiratory Distress Syndrome
Abstract
Background: Acute Respiratory Distress Syndrome (ARDS) is a critical illness syndrome affecting casualties and patients with trauma, sepsis, pneumonia, and aspiration, and it has a 35% mortality rate. The main problem in ARDS is that fluid leaks into the lungs, which makes breathing difficult or impossible for patients. For care providers to appropriately identify patients with ARDS, they must accurately interpret the patient s chest X-ray. Unfortunately, physicians vary widely in how they interpret chest X-rays for ARDS, and we have shown that this varied interpretation is a key factor in the difficulty that physicians have identifying ARDS. Multiple studies demonstrate that up to 65% of patients with ARDS are recognized late or completely missed, and because of this they do not receive evidence-based treatments that improve patient outcomes. Thus, there is an urgent need to develop computational tools that can analyze chest radiology studies automatically to accurately identify ARDS, which would help eliminate human variation and error for this deadly illness. Objectives and Rationale: In this proposal, we will use novel approaches from the field of computer vision to develop deep neural networks (a deep learning methodology) that can identify ARDS on digital chest X-ray images. Computer vision is a field of artificial intelligence (AI) focused on training computers to interpret and understand visual images with human-level accuracy. To accomplish this goal, we will use a large sample of publicly available chest X-rays and a unique chest X-ray dataset from a large cohort of patients from the University of Michigan to train these models. We will also perform two rigorous validation studies of the ARDS network we develop. We will test the performance of the developed ARDS network against individual physician experts and at an external center. This validation procedure will help us to ensure that the technology we develop provides expert-level performance and is generalizable to other patients and health systems outside our own university. Focus Areas: The proposed technology will be a major advancement in detection and care of patients with ARDS. This proposal is relevant to two AIMM FY19 Focus Areas: (1) Algorithm development for decision support in a deployed or operational environment, and (2) AI/deep learning for analyzing and interrogating large medical datasets to identify patterns/predictors of disease. Specifically, we will be using our deep learning methods to analyze large datasets of chest X-ray images, and from this we will develop models that help identify patients with ARDS based on their chest X-ray. Innovation and Clinical Impact: By providing rapid, automated radiologic interpretation of chest X-rays, this technology will be a fundamental leap forward for ARDS care, and it will address a critical limitation in the current diagnosis of ARDS. Notably, we expect that our computer vision network will have better performance than individual physician experts. We will also develop a novel network output display, which will allow care providers to better interpret and evaluate the network s result; this is critical for such a high-stakes setting. Benefit to Service Members, Veterans, Military Beneficiaries, and the American Public: We anticipate that our proposed technology will ultimately be deployed on portable digital imaging systems, enabling real-time processing and analysis of chest X-ray images at the point of care. Because of this, the technology could provide automated radiologic interpretation deployable to forward operating surgical centers as well as advanced echelons of care. This would provide critical diagnostics for military casualties, alerting front-line providers about developing ARDS and enabling rapid identification and triage of ARDS patients to ensure prompt treatment. Our technology will also be extremely useful in civilian hospitals where ARDS is often missed
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 15, 2021
- Source ID
- W81XWH2010496
Entities
People
- Michael Sjoding
Organizations
- United States Army
- University of Michigan