Creating a Statistical Atlas of Prostate Cancer for Guiding Prostate Biopsy

Prostate cancer is the leading cause of death for American men. Transrectal Ultrasonography (TRUS) guided systematic needle biopsy of the prostate has been widely used clinically for the diagnosis and staging of prostate carcinoma. However, due to the limitations of the accuracy of current biopsy procedures, a significant number of prostate cancer cases remain undetected at their initial biopsy. While some researchers have investigated techniques for establishing a 3D distribution map of prostate cancer, the accuracy is limited by the considerable inter-individual morphological variability. This problem can be overcome via elastically deformable anatomical models, which spatially normalize the prostate images to a canonical coordinate system. With accurate registration of the prostate, both the statistical atlas of cancer distribution and the conditional probabilities among cancers in different regions can be obtained and further applied to suggest optimal biopsy strategies. The overall goal of the proposed project is to develop deformation techniques for creating a statistical atlas of cancer and adapting it to the individual morphology of patients by combining knowledge from prostate ultrasound images and shape statistics on training samples.


• Prostate model.




• A subject warped to the model. Overlay of the prostate boundary of the subject (red) after deformable registration with the template prostate (white).

Click here for a deformation movie


• The middle cross-sections of ten prostate subjects, before and after normalization. The red regions denote prostate cancer.




• Optimal biopsy strategy using statistical atlas of cancer distribution. We tested our needle optimization method on 100 subjects. In the following figure, the optimal biopsy sites are shown as white spheres and the prostate capsule is shown as red. The underlying spatial statistical distribution of cancer inside of prostate capsule is shown as green. Brighter green indicates higher likelihood in finding cancer in that location. In this case, seven needles were adequate to detect the tumor 100%, in those 100 subjects. In particular, the first five needles can detect the tumor 97%. An important implication is that the optimized needle placement is not necessarily on regions that have high likelihood of cancer. As we can see from the following figure, only first three white needles were placed in brighter green (high likelihood) regions. The remaining four were placed in regions that were almost statistically independent from the first three.