CRII: SCH: Cardiomegaly detection from deep learning to clinical trials

Cardiac disease is one of the leading causes of death in both humans and animals. Clinicians usually manually analyze cardiac images, which is time-consuming, tedious, error-prone, and expensive. Therefore, automatic detection of enlargement of the heart (i.e., cardiomegaly) is significant in helping clinicians improve diagnosis accuracy.

Deep learning models have been widely applied in image classification tasks of animal cardiomegaly detection; however, clinicians with less background in deep learning lack trust in the results. On the one hand, clinicians are used to detecting the disease using traditional medical metrics (e.g., vertebral heart scale); on the other hand, deep learning models still lack explanations for the relationship between prediction results and input images.

Therefore, a major goal of this project is to bridge the clinical trials' frequently used metrics with deep learning models to improve the trustworthiness of the automated quantification of cardiomegaly disease, leading to the improvement of cardiomegaly diagnosis effectiveness. The project aims to develop novel deep-learning models that yield understandable and matched prediction results with clinical metrics.

Ultimately, this project will build a unified human-computer interaction software for cardiomegaly labeling, detection, and report generation for the general public without prior domain knowledge. Moreover, this project will support the research of Ph.D., Master's, and Undergraduate students and the development of a graduate-level capstone course at Yeshiva University.

This project aims to achieve (1). the development of novel principles and methods for animal cardiomegaly detection: (a) ensuring the perpendicularity of traditional clinical metric by designing a novel perpendicular fully connected layer (PFCL), which performs better than conventional FCL. (b) creating an end-to-end deep detection model to predict biomarkers' positions and present trustworthy cardiomegaly detection results by increasing the accuracy of the vertebral heart scale calculation; (2). automatic cardiomegaly report generation with limited training samples for initial diagnosis by finding semantic mappings between radiographs and the generated texts using deep semantic mapping method and few-shot generation techniques; and (3). designing a novel interface to support the combination of tasks, including data labeling, results prediction, report generation, and modification. These three research goals will be complemented by developing a set of methods to evaluate cardiomegaly detection and interface from users' perspectives.

Clinicians and the general public will be able to use the high-precision open-source animal cardiomegaly diagnosis software to increase diagnosis accuracy, lower diagnosis costs, and reduce emotional stress, especially for pet owners.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.