April 10, 2018

A bit more on semantic segmentation, now 3D

{V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation


--> Link / authors arxiv.org/abs/1606.04797, Fausto Milletari / Nassir Navab / Seyed-Ahmad Ahmadi

--> Essence:

(0) Essentially applies UNet to 3D with a custom DICE based loss

(1) Architecture - goo.gl/Yn2BGb - basically UNet with 3D convolutions. Upsampling / downsampling - goo.gl/VtXrXy

(2) PReLu (no ablation test)

(3) Receptive fields of layers - goo.gl/FGwDCF

(4) 3D DICE loss - goo.gl/SqrK93 (wo BCE?)

--> The paper does not use all the juice possible - hacky transfer learning (obvious idea - just stacking Imagenet filters), CLR, LinkNet architectures, etc

--> Looks like a good baseline / reference

{An application of cascaded 3D fully convolutional networks for medical image segmentation


--> arxiv.org/abs/1803.05431, a group of Japanese researchers

--> Essence:

(0) 2 stage 3D UNet, ablation test against 2D FCNs

(1) Loss - 3D cross-entropy

(2) Transfer learning - it works for other datasets, give a mild boost (1-3 %)

(3) 80-90% DICE, varies by organ

(4) weights downloadable github.com/holgerroth/3Dunet_abdomen_cascade (Caffe...)

--> Essentially a 2 stage process is dictated by memory considerations:

(0) Pipeline goo.gl/wZwF3X

In the long run transfer learning may rule, but here legal limitations may slow down this process.