A versatile method for bladder segmentation in computed tomography two-dimensional images under adverse conditions.

This article presents the design and evaluation of an algorithm for urinary bladder segmentation in medical images, from contrastless computed tomography studies of patients suffering from bladder wall tumours. These situations require versatile methods of segmentation, able to adapt to the structural changes the tumours provoke in the bladder wall, reflected as irregularities on the images obtained, creating adversities to the segmentation process. This semi-automatic method uses fuzzy c-means clustering, a Gaussian-curve-based intensity transformation, and active contour models, requiring only the physician's input of a single seed point for each anatomical view, in order to segment the bladder volume in all frames that include it. The performance of the method was evaluated on eight patients of The Cancer Genome Atlas-Urothelial Bladder Carcinoma collection, achieving approximately 79% of successful segmentations for small tumour patients (below 2.0 cm of diameter) and approximately 72% between 2.0 and 2.9 cm. Successful segmentations for small tumour patients presented an average of 3.7 mm Hausdorff distance and 91.0% degree of overlap. The promising performance attained, especially for small tumour patients, revealed a high potential of this method to serve as basis for an effective early-stage bladder wall tumour computer-aided diagnosis system.