Candy Cane: Breast Cancer Pixel-Wise Labeling with Fully Convolutional Densenets

Breast cancer is one of the leading cancer-related death causes worldwide. Analysis of histology whole-slide and microscopy images is essential for early diagnosis. However this task is time consuming and has to be performed by specialists. Computer-aided Diagnosis can accelerate the process and reduce cost. Candy Cane, a system for performing pixel-wise labelling of whole-slide images in four classes: normal tissue, benign lesion, in situ carcinoma and invasive carcinoma is introduced. It uses fully convolutional Densenet architecture which performs successive convolutions and downsampling on input image followed by successive upsampling and transpose convolution to predict the label image. This allows the system to examine both microscopic low level details as well as high level features and organization of the tissue to make its pixel labeling decision. Accuracy s score of 0.501 are achieved in four-class pixelwise labeling of images on test set.