Recognizing Microscopic Structures: Dense Semantic Segmentation of Multiple Histopathological Classes using Fully Convolutional Neural Networks

Abstract: In order to alleviate the financial burden on the healthcare sector as well as relax its employees’ workload, there is a need to introduce novel tools that automate some of the tasks that today are performed manually. Especially pathology poses a problem with few pathologists, demanding manual labour and unnecessary work on benign tissue. As a response, the DOGS project aims to develop a tool to automate or assist in Gleason grading of histopathological images from prostate biopsies. It is probable that such a tool would benefit from having access to individually segmented, pathologically relevant objects from the images. Moreover, considering recent advances in deep learning and its frequently impressive performance on various image analysis tasks, it is natural to approach this challenge from a deep learning perspective. This thesis proposes several fully convolutional neural networks to be used for dense semantic segmentation on histopathological images. The networks’ architectures are all initially based on already proven networks but are modified in various ways to achieve better performance. Being a supervised machine learning task, the ground truth required to train the network has been developed as a part of the thesis. The best-performing network obtained an accuracy of 79.71 % mean intersection over union and the networks presented plausibly equaled or outperformed state-of-the-art methods in nuclei segmentation. However, further work is deemed necessary for reaching adequate segmentation performance. Several suggestions for possible future directions of work are presented, as well as obstacles that have to be considered moving onwards.