Can AI Identify Cancer Subtypes from Pathology Images?

Author: Rasit Dinc

Introduction

The classification of cancer has traditionally relied on the microscopic examination of tissue samples by pathologists. This morphological analysis, while foundational to oncology, is not without its limitations. Sub-optimal reproducibility in clinicopathological parameters, especially in high-grade tumors, can lead to inaccuracies in assessing the risk of disease recurrence and death, potentially leading to the over- or under-treatment of patients [1]. The advent of molecular classification has provided a more objective and reproducible framework, but even within these defined categories, significant clinical and prognostic heterogeneity persists. This is particularly evident in the "No Specific Molecular Profile" (NSMP) subtype of endometrial cancer, which accounts for approximately 50% of cases and encompasses a wide spectrum of clinical outcomes [1, 2].

In recent years, the convergence of digital pathology and artificial intelligence (AI) has opened up new frontiers in cancer diagnostics. AI, particularly deep learning (DL) models like convolutional neural networks (CNNs), has demonstrated a remarkable ability to analyze high-resolution digital pathology images, or whole-slide images (WSIs), and extract complex patterns that may be imperceptible to the human eye. This has raised a critical question for the medical community: Can AI move beyond simple tumor detection and reliably identify distinct cancer subtypes from pathology images alone? The evidence increasingly suggests that the answer is a resounding yes.

The Power of AI in Deciphering Tumor Morphology

AI-driven analysis of pathology images is not merely about automating the work of a pathologist; it is about augmenting their capabilities with a new layer of quantitative, data-driven insight. These algorithms translate the visual data a pathologist studies into numerical data that a computer can analyze at a massive scale. By training on vast datasets of annotated images, DL models learn to recognize the subtle morphological signatures associated with different cancer subtypes.

A groundbreaking study published in Nature Communications demonstrated the power of this approach in endometrial cancer (EC). Researchers developed a deep learning-based image classifier to differentiate between the NSMP and the more aggressive p53 abnormal (p53abn) subtypes. The AI model not only accurately distinguished between the two but also identified a previously unrecognized subgroup within the NSMP category, which they termed "p53abn-like NSMP." This subgroup, while molecularly classified as NSMP, exhibited morphological features similar to p53abn tumors and, critically, was associated with markedly inferior progression-free and disease-specific survival. This finding, validated across three independent cohorts, showcases AI's potential to refine risk stratification beyond conventional molecular and pathologic criteria [1].

Enhancing Diagnostic Accuracy and Accessibility

The application of AI in classifying cancer subtypes extends to other malignancies as well. In the challenging field of pediatric sarcomas, which are rare and highly diverse, AI models have shown impressive accuracy. A study presented at the American Association for Cancer Research (AACR) Annual Meeting reported that AI algorithms could distinguish between various sarcoma subtypes with accuracy rates ranging from 87.3% to 95.1% using only routine pathology images [3]. For instance, the model differentiated Ewing sarcoma from other types with 92.2% accuracy and could distinguish between alveolar and embryonal rhabdomyosarcoma with 95.1% accuracy [3].

These capabilities have profound implications for clinical practice. Firstly, AI offers a path to greater diagnostic consistency and objectivity, reducing the inter-observer variability that can occur among pathologists. Secondly, and perhaps more importantly, it promises to democratize expertise. The diagnosis of rare or complex cancers often requires review by highly specialized pathologists, a resource not available in many healthcare settings. AI-based models, which can be designed to run on standard computer hardware, could provide clinicians in under-resourced environments with access to highly accurate, streamlined diagnostic support, helping to guide and optimize treatment for more patients regardless of their geographic location [3].

Challenges and the Road Ahead

Despite the immense promise, the widespread clinical implementation of AI for cancer subtyping faces several hurdles. A primary challenge is the availability of large, high-quality, and diverse datasets for training. For rare cancers, assembling a sufficient number of cases is a significant undertaking. Furthermore, variations in slide preparation, staining, and scanning across different institutions can introduce technical artifacts that may bias the model's performance. Therefore, robust data harmonization techniques are essential [3].

Another critical consideration is the potential for algorithmic bias. Studies have shown that some AI models may not perform equally well across different demographic groups, a discrepancy often stemming from imbalances in the training data. Ensuring that these diagnostic tools are equitable and effective for all patients is a paramount concern that requires careful validation and ongoing monitoring.

Conclusion

Artificial intelligence is rapidly evolving from a promising research concept into a powerful clinical tool for cancer diagnostics. By analyzing the intricate details of pathology images, AI models are proving capable of identifying cancer subtypes with a high degree of accuracy, uncovering new prognostic biomarkers, and refining risk stratification in ways that were not previously possible. While challenges related to data acquisition, standardization, and algorithmic bias remain, the potential of AI to enhance diagnostic accuracy, improve treatment optimization, and increase access to expert-level analysis is undeniable. As these technologies continue to mature and integrate into clinical workflows, they are set to become an indispensable partner to pathologists, heralding a new era of precision oncology.

References

[1] Darbandsari, A., Farahani, H., Asadi, M. et al. AI-based histopathology image analysis reveals a distinct subset of endometrial cancers. Nat Commun 15, 4973 (2024). https://doi.org/10.1038/s41467-024-49017-2

[2] Batool, Z., Kamal, M.A. & Shen, B. Advancements in triple-negative breast cancer sub-typing, diagnosis and treatment with assistance of artificial intelligence: a focused review. J Cancer Res Clin Oncol (2024). https://doi.org/10.1007/s00432-024-05903-2

[3] AI-driven Analysis of Digital Pathology Images May Improve Pediatric Sarcoma Subtyping. AACR News Release. (2025). https://www.aacr.org/about-the-aacr/newsroom/news-releases/ai-driven-analysis-of-digital-pathology-images-may-improve-pediatric-sarcoma-subtyping/