Artificial Intelligence for Diagnostic Support in Neuropathology

Qucosa (Saxon State and University Library Dresden)

Abstract

Abstract: The digitisation of histopathology, and the increasing use of whole slide imaging (WSI) have enabled the integration of artificial intelligence (AI) into diagnostic pathology. In parallel, advances in deep learning (DL) and large language models (LLMs) have created new opportunities for computational support across both image- and text-based aspects of medical diagnostics. Neuropathology presents a compelling opportunity for the use of such technologies due to its morphological complexity and evolving classification frameworks, particularly the integration of molecular markers into the diagnosis of adult-type diffuse gliomas in the 2021 WHO classification of central nervous system tumours. This thesis is based on two original studies conducted between 2023 and 2024, and a subsequent Letter to the Editor published in 2025. This work investigates the potential of AI as clinical decision-support tools in neuropathology through two complementary approaches: 1. Deep learning models applied to histological WSIs to predict molecular alterations and tumour subtypes. 2. Large language models applied to free-text neuropathology reports to support interpretation and diagnostic decision-making. In the first study, convolutional neural networks (CNNs) using weakly supervised attention multiple instance learning (attMIL) were trained on haematoxylin and eosin (H&E) WSIs to predict the key molecular features of adult-type diffuse gliomas, including IDH mutation status, 1p/19q codeletion, and ATRX loss. The models achieved strong internal performance, with AUROC values up to 0.95 for IDH mutation prediction and glioblastoma subtype stratification. External validation demonstrated a decline in performance, underscoring the challenges posed by dataset heterogeneity, institutional variability, and highlighting the need for robust multi-centre validation prior to clinical implementation. Publication II evaluated the ability of three contemporary LLMs to interpret descriptive neuropathology text and generate diagnostic outputs. When equipped with in-context learning, the models demonstrated high concordance with expert ground truth, reaching alignment rates of up to 90%. Performance varied between models and prompting strategy, and occasional hallucinations and diagnostic inaccuracies highlighted the need for expert oversight and structured validation frameworks. Publication III forms part of a subsequent scholarly correspondence following Publication II. It addressed methodological and terminological considerations related to in-context learning and retrieval-based approaches, thereby contributing to conceptual clarity in the emerging application of LLMs in pathology. Collectively, this thesis demonstrates the feasibility of applying AI to both the visual and textual aspects of neuropathological diagnosis. The findings support the view that AI can augment, rather than replace, expert interpretation by enhancing efficiency and providing decision support. At the same time, this work identifies critical barriers to clinical translation, including generalisability, regulatory oversight, algorithmic bias and data governance. The integration of AI tools into neuropathology will require not only technical and regulatory development but also adaptation within clinical education and routine practice. Pathologists will need to acquire new competencies in digital literacy, critical appraisal of algorithms, and interdisciplinary collaboration. By addressing both methodological performance and translational challenges, this thesis contributes to the development of a framework for responsible and clinically meaningful AI deployment in diagnostic neuropathology.:Table of Contents / Page Number Dedication / 5 Acknowledgements / 7 Table of Contents / 9 List of Tables and Figures / 11 Abbreviations and Definitions / 13 English Abstract / 17 Abstract auf Deutsch / 19 Introduction to Pathology / 22 Tissue Processing / 22 Immunohistochemistry and Molecular Analysis / 24 The Diagnostic Process: General Pathology Workflow / 26 The Diagnostic Approach for Gliomas / 28 Introduction to Digital Pathology / 34 Introduction to Computational Pathology / 38 Artificial Intelligence, Machine Learning and Deep Learning / 40 A Typical DL Pipeline / 43 Transformers and Large Language Models / 46 From Research to Clinical Application in Pathology / 48 Current Status and Challenges in Neuropathology / 52 Thesis Scope / 54 Results - Publications / 56 I: Direct Image to Subtype Prediction for Brain Tumours / 57 II: Large Language Models as a Diagnostic Support Tool in Neuropathology / 58 III: Authors' reply / 59 Discussion / 61 Broader Impact / 64 Future Outlook / 65 Conclusion / 67 Summary / 69 Background / 69 Research Aim / 70 Methods and Results / 70 Discussion / 73 Conclusion / 74 Zusammenfassung / 75 Hintergrund / 75 Zielsetzung / 76 Methoden und Ergebnisse / 76 Diskussion / 79 Schlussfolgerung / 80 References / 83 Declarations / 93 Annex 1 / 93 Annex 2 / 95 Impact Factor Report / 96

Authors 1

1. Katherine J. Hewitt corresponding Aachen Department of Medicine III

   RWTH Aachen University · Universitätsklinikum Aachen

   Affiliation as printed

   Department of Medicine III , University Hospital RWTH Aachen , Aachen , North Rhine-Westphalia , Germany