Gland Segmentation in Colon Histology Images by FIMM Team Dmitrii Bychkov and Riku Turkki dmitrii.bychkov@helsinki.fi and riku.turkki@helsinki.fi University of Helsinki / Institute for Molecular Medicine Finland Clinical Informatics and Image-based Diagnostics Group Motivation We have been seeking for a model which: Is fairly simple to allow for end-to-end training Does not require additional pre/post-processing Learns the features from the data Network Structure RGB input image 7x7 convolution nxm @3 2 x 2 max pooling nxm @ 64 11 x 11 convolution n/2 x 2m/2 @ 64 2 x 2 max pooling n/2 x m/2 @ 64 7x7 convolution n/4 x m/4 @ 64 5x5 convolutional up-scaling n/4 x m/4 @ 128 Convolutional Neural Network Architecture 3x3 convolutional up-scaling n/2 x m/2 @128 nxm @1 Inference RGB input Decision value map Binary mask Training Scores: f(x; W)(I,j) Cost: 1 J= G 1 ∑ max{0, 0.5− f (x;W )i, j } + B ∑ max{0, 0.5+ f (x;W )i, j } + λ R(W ) i∈G i∈B Errors over epochs: Results Summary The proposed approach: can analyse an image in ~ 1 min achieved near 90% pixel level agreement But it: yields somewhat noisy predictions struggles with high-grade samples requires fixed-size input