Artifact Detection for Reliable Computational Pathology Systems using Artificial Intelligence

Abstract

Cancer is a significant global health concern, resulting in millions of deaths worldwide each year. Cancer diagnosis is performed using histological examination of glass slides under a microscope. Thanks to digitization in recent years, histological slides are scanned and stored as whole slide images (WSIs). These WSIs play a crucial role in the research and development of computational pathology (CPATH) systems for assisting pathologists by automating diagnosis, prognosis, extracting regions of interest (RoIs), visualization, etc. However, the acquisition and digitization of histological slides, during various laboratory stages, introduce artifacts and variations in WSI. Artifacts add irrelevant morphological features that do not contain any cancer-related information. In addition, WSIs collected from different laboratories or scanning hardware may exhibit vast differences in color appearance due to the cancer type and biopsy, age of the slide, file formats, etc. Pathologists ignore these variations and artifacts during manual inspection. However, CPATH systems depend on the WSIs and may include them as noise in the learning or prediction process. CPATH systems using artificial intelligence (AI) can unfold information embedded in WSIs. However, artifacts can significantly impact CPATH systems and result in unreliable predictions. Therefore, it is essential to equip CPATH systems with artifact processing pipelines to detect and discard artifacts before running a diagnostic algorithm. An effective pre processing pipeline will enhance CPATH systems’ performance by ensuring diagnostically relevant areas and serve as a quality control tool. This thesis provides a comprehensive review of WSI preprocessing, focusing on the process of WSI acquisition and linking every step to the appearance of artifacts. This thesis highlights the need for further research in artifact detection through deep learning (DL) techniques. In our thesis, we propose novel DL methods to detect notable artifacts and assess the im pact of color processing on detecting artifacts using state-of-the-art (SOTA) architectures. The DL models are trained using various learning strategies such as transfer learning (TL), knowledge distillation (KD), uncertainty aware deep kernel learning (DKL), and the mixture of experts (MoE) scheme, with the aim of creating more robust and generalized methods. Finally, we develop end-to-end artifact processing pipelines to extract, detect, and eliminate artifacts from WSIs and perform quality control. The proposed artifact processing pipeline provides segmentation masks, highlights potential diagnostically relevant areas, generates artifact reports with a percentage of artifact-free regions, and refines artifacts from WSI for further analysis. Moreover, we highlight the privacy challenges and progress in histological data sharing for developing DL algorithms and CPATH services. The methods (TL, KD, DKL, MoE) proposed in this thesis outperform SOTA techniques for artifact detection and generalize well with little training data. Our contributions aim to take a step towards robust and reliable CPATH systems, impacting digital pathology workflow in practice. The source codes and the training and development dataset are made publicly available as open-science contributions by this thesis.