3D detection and extraction of bladder tumors via MR virtual cystoscopy

This paper proposes a pipeline for the detection and extraction of 3D regions of bladder tumors via MR virtual cystoscopy.

After the acquisition of volumetric bladder images with a high-resolution T2-weighted 3D sequence, the inner and outer surfaces of the bladder wall were segmented simultaneously by a coupled directional level-set method. Based on the Laplacian method, a potential field was built up between two surfaces so that the thickness of each voxel within the bladder wall was estimated. To detect bladder abnormalities, four volume-based morphological features, including bent rate, shape index, wall thickness, and a novel morphological feature, which reflects bent rate difference between the inner and outer surfaces, were extracted. The combination of these four features was used to detect seeds on the inner surface by using selected filtering criterion. Then all points on streamlines started from detected seeds formed 3D candidate regions. Finally the fuzzy c-means clustering with spatial information (sFCM) was used to extract tumors from surrounding bladder wall tissues in candidate regions.

The proposed pipeline was evaluated by a database of MR bladder images acquired from ten patients with bladder cancer. To find an optimal feature combination for tumor detection, the performance of different combinations of these features was evaluated with different filtering criteria. With the combination of all four features, the computer-aided detection pipeline shows a high performance of 100 % sensitivity with 2.3 FPs/case. Comparing with tumor regions delineated by radiological experts, the average overlap ratio of tumor regions extracted by sFCM is 86.3 %.

The experimental result demonstrates the feasibility of the proposed pipeline on the detection and extraction of bladder tumors. It may provide an effective way to achieve the goal of evaluating the whole bladder for tumor detection and local staging.