Meeting Banner
Abstract #2399

Evaluation of convolution strategies for 3D-CNN for MR Prostate segmentation

Yanlu Wang1,2, Patrick Masaba3, Fredrik Strand4,5, and Fredrik Järderling6,7
1Department of Clinical Sciences, Intervention and Technology, Karolinska Institute, Stockholm, Sweden, 2Medical Radiation Physics and Nuclear Medicine, Karolinska University Hospital, Solna, Sweden, 3Imaging and Function, Karolinska University Hospital, Stockholm, Sweden, 4Breast Radiology, Karolinska University Hospital, Solna, Sweden, 5Department of Oncology-Pathology, Karolinska Institute, Stockholm, Sweden, 6Department of Molecular Medicine and Surgery, Karolinska Institute, Stockholm, Sweden, 7Department of Radiology, Capio S:t Göran Hospital, Stockholm, Sweden

Typical prostate MR dataset are 2D with few slices and high in-plane resolution. These dimensions do not conform naturally to traditional 3D convolution neural network up/down-sampling schemes, which customarily double/halves the input dimensions in all directions. We compared two strategies for up/down-sampling applied to prostate segmentation in MRI datasets: 1) Transform the datasets and applying orthodox convolution layers. 2) Adapt the up/down-sampling convolution layers in order to use more native resolution input datasets. Our results show that, in general, the former strategy works best, both in terms training loss convergence and overall segmentation results.

This abstract and the presentation materials are available to 2020 meeting attendees and eLibrary customers only; a login is required.

Join Here