Assessing Mammography AI

While routine screening has been shown to decrease breast cancer mortality in multiple randomized controlled trials, mammography is limited by subjective human interpretation. Recent advances in improved computer processing power, cloud-based data storage capabilities, and availability of large imaging datasets have led to renewed excitement for applying artificial intelligence (AI) to mammography interpretation. A collaboration between the University of Washington (Christoph Lee) and UCLA (Joann Elmore, Arash Naeim, William Hsu), this project is undertaking a unique academic-industry partnership to validate, refine, scale, and clinically translate our proven 2D mammography AI algorithm to 3D mammography interpretation. We are validating an existing algorithm for 2D mammography using UCLA’s Athena Breast Health Network. We are enhancing this 2D AI algorithm with expert radiologist supervision and will be examining the impact of adding novel non-imaging data parameters, including genetic mutation and tumor molecular subtype data, to help train the AI model to identify more clinically significant cancers.

We are using several novel technical algorithmic approaches to scale from 2D to 3D mammography that in our preliminary studies have shown improved accuracy beyond radiologist interpretation alone. A series of interpretive studies are planned to identify the optimal interface between “black box” outputs and radiologist interpreters, which remains an understudied topic. Our new end-user tool aims to help tip the balance of routine screening towards greater benefits than harm.

Improving Prostate Cancer Diagnosis

Prostate cancer is the second leading cause of cancer death in American men, accounting for 26% of new cancer diagnoses and 9% of cancer deaths in men. Active surveillance, radical prostatectomy and radiotherapy are commonly used treatments for clinically localized prostate cancer. However, current risk stratification methods cannot be used effectively to avoid subjecting patients with clinically indolent cancers to unnecessary interventions, causing significant morbidity and cost.

The research objective of this R21 is to develop novel techniques using multiparametric magnetic resonance imaging (mp-MRI) and MRI-ultrasound (US) fusion guided biopsy data that provide discriminatory power in distinguishing indolent versus clinically significant prostatic adenocarcinoma based on non-invasive imaging. We are implementing a multi-instance learning (MIL) based convolutional neural network (CNN) model for clinical prostate mp-MRI sequences to generate new quantitative imaging features representative of the underlying tissue. Our MIL-CNN model accomodates ground truth labels from pathology whole mount specimens, as well as MRI-US fusion biopsy results. Hierarchical CNN features will be used to predict voxel-level cancer suspicion, thereby enabling a novel method for performing “imaging biopsy.” Finally, voxel-level suspicion maps will be aggregated into patient-level quantitative imaging biomarkers and combined with clinical data to create a multimodal nomogram for performing risk stratification.

CDS for Indeterminate Pulmonary Nodules

The rollout of low-dose computed tomography (LDCT) lung screening programs is accelerating in the United States, aiming for earlier detection of lung cancer to improve long-term survival. However, a consequence of such imaging programs is the increased discovery of indeterminate pulmonary nodules (IPNs). Significant questions remain around the effective management of screen- and incidentally-detected IPNs: while many are benign, a fraction will go on to become cancerous.

The objective of this R01 is the development of a clinical decision support tool for the management of screen- and incidentally-detected IPNs. We address two key challenges: 1) the development of a continuous-time model for predicting how the IPN will evolve; and 2) the use of this prediction to determine a series of actions over time that will optimize (screening) outcomes for the individual. We first explore the development of a continuous time belief network (CTBN), a temporal probabilistic model to predict the likelihood of a patient to develop lung cancer. The probabilities computed through the CTBN are subsequently input into a partially-observable Markov decision process (POMDP) to guide IPN management decisions.

Understanding Stroke

Stroke is a leading cause of mortality and morbidity in the United States, with approximately 795,000 Americans experiencing a new or recurrent stroke each year. Intravenous tissue plasminogen activator (IV tPA) is the dominant and most proven treatment option, but its use is only indicated within 4.5 hours following a stroke. Unfortunately, up to 30% of stroke patients present with an unknown time since stroke (TSS) symptom onset, which makes them ineligible to receive IV tPA. Many of these individuals could be spared severe morbidity or mortality if there existed an alternative method for establishing TSS, allowing them to be identified and treated.

In this project, we aim to: 1) develop a machine learning framework for classifying TSS; 2) develop a deep convolutional autoencoder to generate novel multimodal image representations from MR and CT to improve classification; and 3) implement visualization techniques that elucidate the relationship between deep features and pathophysiological stroke processes.