Robust and Interpretable Multimodal Machine Learning Models for Diagnosis and Prognosis of Prostate Cancer

ABSTRACT Prostate cancer is the second leading cause of cancer death and affects about 13 out of 100 American men. The standard diagnosis approach for prostate cancer incorporates prostate specific- antigen (PSA) blood test, digital rectal examination and biopsy, and, recently, the Prostate Imaging Reporting and Data System (PI-RADS) that localizes and stratifies the risk of lesions in biparametric

MRI (bpMRI) images. Detection of clinically significant cancer (csPCa) remains a challenge, resulting in over- and underdiagnosis. Thus, accurate diagnosis and staging of prostate cancer will have a significant impact on successful treatment planning and clinical management. As large-scale imaging datasets become widely available in clinical settings, more prostate cancer studies are adopting

machine learning (ML) and deep learning (DL) techniques. However, the common approach in the existing studies has a drawback of using imaging findings, e.g., lesion size and radiomics features. This project aims to implement and validate cutting-edge DL-based frameworks for the diagnosis and prognosis of prostate cancer while fully utilizing imaging data and clinical scores (e.g., initial PSA

level and prostate genomic score). The proposed project will use state-of-the-art DL architectures such as U-Net, image transformer and neural additive model to develop frameworks for robust DL- based prostate and zonal anatomy segmentation (Aim 1), prediction of a csPCa probability map (Aim 2), and forecast of biochemical recurrence using interpretable multimodal DL framework (Aim 3). We

expect these frameworks will offer novel techniques adapted for prostate cancer, unique and interpretable features from multimodal markers to better understand diagnosis and prognosis, and will eventually help guide treatment selection and clinical management. This project will build on Dr. Kim's quantitative background in modeling and analysis of neuroimaging data. During the project, Dr. Kim

will gain clinical expertise in prostate cancer via formal training driven by coursework (e.g., oncology and biology) and research mentorship. Through the K25 award, Dr. Kim will lay the groundwork for establishing an independent research program of computational methods for prostate cancer.