Automated Deep Learning in Toxicologic Pathology Slide Image Analysis for Pharmaceutical Drug Development

“Automated Deep Learning Toxicologic Pathology Slide Image Analysis for Pharmaceutical Drug Development," investigates the integration of artificial intelligence (AI) and deep learning (DL) to automate the analysis of whole slide images (WSI) in the content of how to scale imaging (and AI) in pharmaceutical drug development with a specific focus on toxicologic pathology stage of the drug pipeline. This research aims to make preclinical safety studies—which are pivotal before commencing human clinical trials—more efficient and accurate, thus accelerating the drug development process. One area where AI systems can dramatically improve WSI analysis is in the context of preclinical toxicology studies for drug development. The generation of hundreds or even thousands of pathology slides in a typical toxicology study presents a significant bottleneck. Against the backdrop of a regulatory environment, the large number of studies underway within a single company or clinical ecosystem must allow for scientific oversight to be applied to govern and ensure both compliance with protocols and safety for patients. As the complexity and volume of WSI grows, there is a greater demand for an innovative solution to manage these individual processing workflows. To meet this challenge, we propose a digital pathology AI processing fabric to facilitate processing, integrating, and developing deep learning and AI models for use at scale in a pharmaceutical company under the supervision of regulatory bodies including the FDA and GxP practices. By automating the initial phases of slide analysis, pathologists can be freed from the monotonous and labor-intensive task of screening slides, allowing them to focus on more complex diagnostic work. This analysis and co-pilot system not only improves efficiency but also ensures consistency and repeatability in slide analysis, which is crucial for regulatory submissions and scientific research. The thesis is systematically structured around three primary aims: • Development of an AI Fabric System: The first aim is the development of an Pharmaceutical Drug Development AI Fabric System designed to automate technical tasks involved in the analysis of WSI, thus providing critical support to pathologists. This system leverages robotic process automation (RPA) and deep learning to streamline the analysis of toxicologic pathology slides. The primary objective here is to digitize and automate large segments of the workflow that pathologists currently perform manually, thereby reducing time and improving accuracy. • Introduction of a Novel Multiclass Weak Labeling Approach: The second aim introduces a novel multiclass weak labeling approach utilizing semantic segmentation for automating the scoring of muscle atrophy. The method uses the DeepLabv3+ architecture, an open-source sophisticated convolutional neural network, to categorize varying levels of atrophy in muscle tissue slides. Traditional strong labeling methods require extensive manual effort from pathologists, which is both time-consuming and subjective. In contrast, the weak labeling approach simplifies this process by allowing an AI model to learn from less precisely labeled data, significantly reducing the burden on pathologists. The study shows that this weak labeling method achieved a concordance of 96.6% with pathologists’ evaluations, outperforming the 75% concordance achieved through traditional strong labeling methods. This finding underlines the efficiency and accuracy benefits of the weak labeling approach. • Practical Deployment within a Pharmaceutical Setting: The third aim addresses the practicality of deploying this system within a pharmaceutical setting, given the stringent regulatory requirements. We explore various deployment strategies, including centralized and decentralized (edge) computing models for processing and storing WSI data. The hybrid approach we propose offers a balance between the robustness of centralized data management and the agility of edge computing. This ensures that data can be handled efficiently while complying with regulatory standards, such as the FDA and GxP guidelines, which require rigorous validation of processes and systems to ensure the integrity and reliability of study results. Furthermore, the thesis underlines that the future work will involve extending this framework to other tissue types and exploring more advanced AI and DL models and how an AI Fabric could help support the heterogeneity in AI methods while unifying it under a centralized platform allowing for regulatory controls to be implemented, while maintaining the agility required for innovation. The continuous evolution of AI technologies means there is immense potential for further improving the accuracy and efficiency of toxicologic pathology analyses. These advancements hold great promise for making toxicologic assessments faster and more reliable, which could significantly shorten the drug development timeline. In conclusion, this thesis presents a comprehensive and forward-thinking approach to integrating AI and deep learning into toxicologic pathology. By automating large segments of the pathology workflow and ensuring regulatory compliance, this research paves the way for faster, more accurate preclinical studies. Moreover, it has the potential impact on the pharmaceutical industry is vast, providing a means to expedite the development of safe, effective drugs and improving the overall efficiency of toxicologic assessments.