Refined Reconstruction of Global Prostate Segmentation from Patch-wise Coarse Predictions

Whole organ segmentation in biomedical images continues to be an important problem. Recently deep learning based methods have produced convincing results. Taking into account limited sample sizes and inconsistent spacing on devices, many solutions are designed to work patch-wise, with networks trained on sub-volumes with a certain stride, rather than on the whole 3D image. The binarized segmentation probability map of the whole volume image is then computed by mapping back stacked sub-volume predictions using a threshold. As may be expected, the performance is highly sensitive to the threshold chosen, as well as issues such as low probabilities on object boundary region, missing voxels or distortions on corners of sliding windows and unexpected components. In this work, we analyse and test the performance of a commonly used thresholding method, and introduce learning based methods to address these issues. In addition, a simple shape prior from a size-based shape heat map is introduced to improve overall performance. Experiments were carried out on the open MICCAI PROMISE12 challenge dataset for prostate segmentation. With the learning based method, the average performance increased by 1.65% on Dice Similarity Coefficient(DSC) compared to the thresholding method. On challenging cases, the improvement was more than 6.55% on DSC. Since proposed method produces convincing results with only modifications of the reconstruction step, this approach may be adopted in other patch-wise deep learning networks too.