Abstract
Artificial intelligence (AI) offers transformative potential in pathology, where histopathological images remain the diagnostic gold standard due to their rich morphological and molecular information. While the rapid development of AI-driven computational pathology tools is revolutionizing disease interpretation, these technologies have not yet been systematically evaluated. Therefore, this review systematically evaluates AI applications across the diagnostic continuum, from image preprocessing and tumor classification to prognostic stratification and the discovery of predictive biomarkers. It presents a technical taxonomy of the algorithms and foundation models powering these applications, benchmarking their performance across diverse diagnostic tasks through rigorous comparative analyses. It also identifies critical challenges in clinical translation, including computational scaling, noisy annotations, interpretability gaps, and domain shifts. Finally, it proposes a roadmap for advancing AI applications in precision oncology and pathological research. By bridging technological innovation with clinical needs, this review aims to accelerate the integration of robust, unified, scalable AI solutions into diagnostic workflows.
Introduction
Pathologists have traditionally diagnosed cancer by manually examining tissue slides under a microscope—a process that is both time-consuming and prone to subjective variability. The digitization of these slides into gigapixel Whole-Slide Images (WSIs) has catalyzed the field of computational pathology. AI now offers the potential to automate and enhance critical tasks such as tumor grading, survival prediction, and the discovery of novel biomarkers. This paper systematically reviews the technological landscape, evaluates current foundation models, and identifies the hurdles preventing real-world clinical adoption.
The Evolution of Computational Pathology
The technological progression of the field can be categorized into three distinct eras:
- Traditional Machine Learning: These methods rely on handcrafted morphological features—such as nuclear shape, texture, and cell arrangement—which are fed into classifiers like Support Vector Machines (SVMs). While highly interpretable, they are limited in their ability to handle complex spatial patterns.
- Deep Learning (DL): Utilizing Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), this era introduced automated feature extraction for segmentation and detection. To manage the massive size of WSIs without requiring exhaustive pixel-level annotations, Multiple Instance Learning (MIL) is utilized, allowing models to learn from high-level slide labels (e.g., "cancer" vs. "normal").
- Foundation Models (FMs): The current frontier involves massive models pre-trained on millions of pathology images using self-supervised learning. Examples include Virchow, UNI, and CHIEF. These models extract highly transferable features that can be fine-tuned for diverse downstream tasks, from rare cancer detection to predicting genetic mutations directly from standard H&E stained slides.
Benchmarking and Data Resources
The development of these tools relies on massive open-source datasets, such as CAMELYON16 & 17 for metastasis detection and The Cancer Genome Atlas (TCGA) for multi-omics integration.
A comparative benchmark conducted by the authors evaluated several foundation models (including HIPT, CTransPath, UNI, and CONCH) across multiple TCGA datasets. The results indicated that the CHIEF model achieved the best overall performance, demonstrating superior diagnostic accuracy (up to 96% in certain cancer types) and outperforming others in prognostic tasks (measured by the C-index) for liver and lung cancers.
Key Insights for Precision Oncology
- The Shift to Weak Supervision: MIL has successfully allowed models to achieve expert-level accuracy using only high-level clinical labels, bypassing the need for expensive cellular-level annotations.
- Multi-omics Integration: Unimodal data is often insufficient to capture tumor complexity. The most predictive models now fuse histopathology with genomics and transcriptomics to provide a holistic view of patient health.
- Zero-Shot Capabilities: Vision-language models (like CONCH and PLIP) enable "zero-shot" classification, allowing them to identify pathology concepts they weren't explicitly trained on by understanding text descriptions in pathology reports.
Challenges and Future Directions
Several critical bottlenecks must be addressed before widespread clinical adoption can occur:
- Computational Scaling: Processing gigapixel WSIs is resource-intensive. Future work must focus on efficient attention mechanisms and data compression to streamline training and inference.
- Domain Shifts and Noise: Varying staining protocols across different laboratories introduce significant domain shifts. Developing noise-resistant training frameworks is essential for model robustness.
- The "Black Box" Problem: A lack of interpretability remains a major barrier to clinical trust. Future models must integrate explainable AI (XAI) techniques to show clinicians exactly why a specific diagnosis or prognosis was generated.
By bridging technological innovation with clinical needs, the integration of robust, unified, and scalable AI solutions will eventually become a standard component of the diagnostic workflow, accelerating the progress of precision oncology.
Summary updated