ISBI Daily - Thursday

the prostate – not only axial scans, but also sagittal (left to right) and coronal (front to back) scans – and processes them in an end-to-end fashion in one neural network. Anneke tells us the volumes usually have very high slice thicknesses and poor resolution. It was not difficult to combine them, but to get a higher resolution output out of them that has isotropic voxel sizes. She explains: “ The high MRI slice thickness for axial prostate volumes becomes problematic especially at the top and bottom region of the prostate. By considering additionally sagittal and coronal volumes, we aim to obtain more reliable results for these region s.” Anneke says that it can be a challenge when you only have limited capacity in your network and GPU. However, with a more powerful GPU , they could get more precise results and would have more data. They currently use a limited public dataset, but it’s only for prostate cancer prediction or assessment – for grading the prostate cancer – and their medical expert did manual segmentations. Anneke would like to be able to evaluate the method on a larger dataset in the future. In terms of next steps, Anneke tells us that sometimes the volumes themselves might not be perfectly aligned if a patient moves during the MRI scan, so they would like to include automatic registration to overcome this problem. She would also like to refine the network further, because it is just a basic architecture that was modified for this purpose. Anneke Meyer ISBI DAILY Thursday 11 To find out more about Anneke’s work, come along to her poster session today at 11:30 in the Ambassador Ballroom. “ By considering additionally sagittal and coronal volumes, we aim to obtain more reliable results for these regions. ”