Computer-Aided Image Analysis and Decision Support System for Bladder Cancer

Abstract

Bladder cancer is a common type of neoplasm that can cause substantial morbidity and mortality among patients. Bladder cancer causes 16,870 deaths per year in the United States. It is expected that 76,030 new bladder cancer cases will be diagnosed in 2017. Multi-detector row CT (MDCT), and specifically MDCT when used for urography (CTU), has become the imaging modality of choice for evaluation of most urinary tract abnormalities. Interpretation of MDCT urograms that commonly exceeds 400 images. This is a demanding task for radiologists who have to visually track the entire upper and lower urinary tract and look for lesions that usually are small in size. Using a computer-aided diagnosis (CAD) system as an adjunct for the radiologist may reduce the number of lesions that are missed by the radiologists. To create a CAD system for detection of bladder lesions, we have developed methods and software to perform the following specific tasks: (1) segment the bladder; (2) detect bladder lesion candidates; (3) segment the bladder lesion candidates; (4) extract features from the segmented lesion candidates; (4) classify lesion candidates as true lesion findings or false positives using the extracted features; (5) detect bladder wall thickenings. Correct staging of bladder cancer is crucial in determining the need for neoadjuvant chemotherapy treatment and minimizing the risk of under-treatment or over-treatment. Also, reliable assessment of the response to neoadjuvant therapy at an early stage is vital for identifying patients who do not respond to this treatment and allows the physician to discontinue ineffective treatment and its undesired adverse effects on a patient’s physical condition. We have developed prototype predictive models for both bladder cancer staging and bladder cancer response to neoadjuvant treatment by adapting the methodology of feature classification that merges image-based biomarkers. The bladder lesion segmentation modules were used to extract the image-based biomarkers as input to the models. Early detection of bladder cancer, accurate tumor staging, and early prediction of treatment response could reduce mortality and morbidity, and improve quality of life for surviving patients. This dissertation presents the methods that we developed to automatically segment the bladder on CTU, and automatically detect masses and wall thickenings with sensitivity near 90% with a relatively low number of false positive findings. We have developed a system that distinguishes between muscle-invasive and non-muscle-invasive cancer, which is a clinical threshold used to determine treatment. We have also developed a system that uses the pre-treatment and post-treatment CTU scans to estimate the likelihood that the patient has fully responded to treatment. This system has achieved performance comparable to the radiologists, with area under the receiver operator characteristic curve (AUC) values in the range of 0.69 to 0.77. We performed an observer performance study where we saw that our system of predicting complete response to treatment improves the performance of the clinicians when they read the pre- and post-treatment scans with the aid of the system. On average, the clinician AUC increased with statistical significance from 0.74 to 0.77. Early detection of bladder cancer, accurate staging of tumors, and early prediction of treatment response can reduce mortality and morbidity, and improve the quality of life for surviving patients. These studies show possible methods for performing those tasks.