Novel deep learning-based method for prostate segmentation in T2-weighted magnetic resonance imaging

Summary

This paper presents a simple but apparently effective method for segmenting prostate MRI images. As illustrated in fig.1, they use 2 CNNs. The first global one roughly segments the prostate. Based on that segmentation, they crop the image and feed it to a second local CNN. The resulting segmentation map is then post-processed with basic morphological operations.

NOTE that their CNNs are UNet are some sort of denseNets with residual modules.

Results

Results on the PROMISE12 challenge dataset seem good.