Deep convolutional network for tumor bud detection

The existence of tumor buds is accepted as a promising biomarker for staging colorectal carcinomas. In the current practice of medicine, these tumor buds are detected by the manual examination of a immunohistochemically (IHC) stained tissue sample under a microscope. This manual examination is time-consuming as well as it may lead to inter-observer variability. In order to obtain fast and reproducible examinations, developing computational solutions has been becoming more and more important. With this motivation, this thesis presents a fully convolutional network design for the purpose of automatic tumor bud detection, for the rst time. This network design extends the U-net architecture by considering up-to-date learning mechanisms. These mechanisms include using residual connections in the encoder path, employing both ELU and ReLU activation functions in di erent layers of the network, training the network with a Tversky loss function, and combining outputs of di erent layers of the decoder path to reconstruct the nal segmentation map. Our experiments on 3295 image tiles taken from 23 whole slide images of IHC stained colorectal carcinomatous samples show that this extended version helps alleviate the vanishing gradient problem and those related with having a high class-imbalance dataset. And as a result, this network design yields better segmentation results compared with those of the two state-of-the-art networks.