Transfer Learning and Data Augmentation for Semantic Segmentation in Histopathology

The segmentation of Region Of Interests (ROIs) in high-resolution histopathological images is a task with high clinical relevance which requires large amounts of reading time from pathologists and suffers from many considerable problems, such as inter and intra-reader variability. Thus, its automation is desirable because it would hugely reduce the workload of the clinicians, while at the same time reduce the subjectivity in diagnosis. However there are still open issues like the amount of data required for training and testing algorithms, the computational load when dealing with high resolution images and, most of all, the large variability in the analyzed data (mainly due to different cell types, cell densities, stains, magnification levels, and so on).The segmentation of ROIs can be expressed as a Semantic Segmentation problem for which Deep Learning models are proven to be particular suitable.However building these models require an huge amount of annotated training data. To overcome this limitation, Transfer Learning and Data Augmentation have been proposed as possible solutions.In this thesis I show how Transfer Learning and Data Augmentation can be leveraged for Semantic Segmentation of histopathological images. To my knowledge, such approach represents a novelty not only in the biomedical field, but also in many other fields of computer vision.In this thesis the UNet convolutional-deconvolutional neural network is implemented in Keras (Tensoflow/Theano backend) and trained using the Camelyon16 Dataset (H&E breast cancer Whole-Slide Images). Whole-Slide Images (WSIs) are first preprocessed (tissue regions segmented from fatty and white background areas and then stain-normalized) and then fed as tiles to the CNN. The outputs are probability maps, whose pixel intensity values express the probability of being part of a cancer metastasis. In a second phase, a fine-tuning of the pre-trained UNet is performed on the GlaS (H&E glands images) and ISBI12 (Drosophila first instar larva ventral nerve cord ssTEM) datasets in order to understand if the trained model can be used for segment completely other types of images for which the available datasets are not sufficient. Fine Tuning was aided by a Data Augmentation pre-processing. Results are very promising and, to my knowledge, this work represents a novelty in the computer vision field.