Development of a deep learning-based method for artifact detection and quality controls in digital pathology

In the last years, in an attempt to overcome some of the issues of traditional pathology, whole slide digital scanners have been adopted, enabling the transition of pathology into the digital era. Histological slides can now be digitalized, but the process that goes from the collection of the tissue to the digital image consists of specific sequential steps, typically carried out manually by laboratory technicians, each of which can introduce artifacts such as tissue folds, air bubble, pen marker or dust, that can lower the quality of the histological image. These artifacts may alter the appearance of the tissue, making diagnosis difficult for the pathologist and negatively affecting the performance of automatic algorithms that operate on digital WSI. For these reasons, a quality control mechanism is needed. Currently, most quality control processes are performed manually, but it is a laborious and subjective task; moreover, there are artifacts that may complicate the quantitative analysis of automated algorithms while not having implications for the pathologist's diagnosis. An automated approach for quality controls can help to overcome these problems. The aim of this thesis is to develop a fully automated quality control system for histological slides. To achieve this goal, a multi-class approach based on deep learning was used which allows, starting from the pyramidal image, to identify the artifacts present on the entire slide. The dataset, including thousands of whole slide images (WSIs) relating to 9 different organs and 4 different stains, was manually annotated, and for each WSI the images at a magnification of 1.25x and 5x and the corresponding manual masks were extracted. After the data preparation, the tiles extracted from the starting images and masks were used to train the neural networks for the segmentation of tissues and artifacts. The architecture used for the neural networks is the DeepLabV3+ model with a ResNeSt as backbone, that is a split-attention network which uses the typical attention mechanism applied on channel to capture cross-channel feature correlations, while preserving a multi-path architecture to learn independent features. For the test phase, a sliding window approach has been used, keeping only the center of each prediction. The results show that this system has a high generalizability. For this task, DSC values higher than 90% were achieved, overcoming the performances of other methods currently present in literature. The approach developed proves to be useful for two main aspects: on one side, introducing an automated quality control system into the daily workflow of pathologists allows to quickly identify poor-quality slides that needs to be reproduced or rescanned, thus speeding up the workflow and avoiding delays in the formulation of diagnoses; on the other side, it can help the development of new automatic algorithms, identifying the regions that should to be avoided during the training and the testing of the algorithm, so as to have more robust and faster algorithms and avoid results distorted by the presence of artifacts. In the future, this model can be further improved by increasing the number of poorly represented organ-color combinations, reducing training times and developing a more structured quality score for histological slides.