Prostate cancer (PCa) is a leading cause of cancer related mortality in men in the United States. Imaging is crucial for PCa staging and treatment decisions. Currently, the gold standard is a multi-parametric magnetic resonance imaging (mpMRI) scan followed by confirmatory ultrasound (US) guided biopsy. While mpMRI has good sensitivity and specificity for index lesion detection, the largest and most aggressive lesion, it is resource intensive, expensive, and generally limited to larger centers. Micro-ultrasound (microUS) is a high-resolution US imaging system that has recently been introduced as a low-cost alternative to mpMRI for imaging PCa. While preliminary results are encouraging, microUS lacks direct comparison with mpMRI and ground-truth whole-mount (WM) pathology. It also requires substantial hands-on experience for accurate interpretation. Ultimately, skepticism relative to established mpMRI imaging and barriers to entry for new users among other reasons have limited microUS adoption. This thesis details a comprehensive investigation into microUS imaging and its role in PCa diagnosis. To this end we performed a clinical trial collecting microUS images of patients undergoing radical prostatectomy for PCa with corresponding mpMRI imaging and WM pathology. Using this dataset, we were able to quantitatively explain the diagnostic ability of microUS, however its utility relative to mpMRI required further investigation. To quantify the diagnostic ability of microUS, expert and novice readers reviewed the collected microUS and mpMRI images. For correlation of the delineated regions of interest (ROIs), we developed and validated a novel co-registration process to allow for direct registration of microUS and mpMRI with WM pathology with a registration error of 3.90±0.11mm. Comparing the confirmed regions of cancer on WM pathology with the identified regions on microUS and mpMRI, we found microUS to have similar detection of index lesions to equivalent MRI review (91.7% vs 80%). In an attempt to alleviate inter-reader variability and the reduced performance of novice users, we trained deep learning classifiers to identify PCa on microUS imaging using our accurately labeled data. These models exhibited expert level performance for image classification and targeted biopsy guidance. Further work is necessary for clinical adoption of deep learning models for microUS imaging of PCa.

Background

Methods

Results

Conclusions