• There are no suggestions because the search field is empty.

Case study: AI-based image analysis enables prognostication in ovarian cancer

A research team from the University of Helsinki used Aiforia® Create to develop a deep learning AI model to predict patient outcomes in a complex ovarian cancer.
Written by Aiforia
Case study: AI-based image analysis enables prognostication in ovarian cancer
7:08

Although high-grade serous carcinoma (HGSC) is the most common ovarian malignancy and the fifth leading overall cause of cancer death in women, little is known about how these tumors initiate their growth and progress into late-stage carcinomas. 

HGSC tumors are aggressive, and they present with significant genetic and morphologic heterogeneity—both between and within individual patients—at a late stage (III-IV), making it challenging to study the underlying disease biology. Whereas studies have attributed long-term survival with HGSC to clinical factors like diagnosis at a younger age or lower disease stage and optimal surgical interventions, patients diagnosed with HGSC still have unpredictable outcomes.

Dr. Anna Laury, a pathologist and a researcher at the University of Helsinki, and her team leveraged Aiforia® Create, Aiforia’s versatile AI development tool, to successfully train a deep learning model to identify morphologic regions that can then help predict patient response to platinum-based therapy in high-grade serous carcinoma.

This proof-of-concept is groundbreaking because it demonstrates that an AI development tool like Aiforia® Create can help pathologists develop and validate deep learning AI models to analyze histological features and patterns, enabling not just the diagnosis but the prognosis of complex, heterogeneous malignancies like carcinomas.

“It is exciting to realize that AI-based image analysis models can identify meaningful morphologic patterns within tumor tissue that are currently invisible to humans.” – Anna Laury, MD, PhD, Clinical Researcher at the University of Helsinki

 

AI can be used for prognosis in pathology

Most deep learning-based AI approaches for whole slide images (WSI) merely automate existing pathology diagnostics, detect metastasis faster, or uncover correlations with known genetic alterations in various cancers. In HGSC, for example, multiple heterogeneous morphologies exist, making it difficult to identify the most informative and relevant tumor regions to study and target with therapies. 

Beyond leaning on histologic slides to confirm a primary carcinoma diagnosis, pathologists don’t typically explore morphological analysis for prognostic information in HGSC, such as to predict a patient’s response to platinum-based therapy. Whereas some studies have attempted to develop predictive tools that can associate specific morphologic features with mutational findings, they have been unable to achieve the sensitivity and specificity necessary to use these tools in daily clinical practice.

As Dr. Laury’s group has shown, deep learning neural networks can identify morphologic regions that predict response to platinum-based therapy in HGSC and associate these morphologies with specific transcriptomic signatures within tumor WSI.

 

Predicting patient survival outcomes using Aiforia® Create

Beyond detecting known features in complex tissue images or precisely quantifying areas in these images, the developed AI model can help pathologists predict survival outcomes of patients with HGSC based on the histological patterns in these tumors. These morphological features were previously unrecognized by human pathologists, meaning the AI model extracts novel, relevant insights to help the pathologists explain these findings. 

When Dr. Laury’s team initially used this AI model’s deep learning neural networks to extract clinically useful information from WSI, they found that these networks could discriminate patient outcome extremes in HGSC using only tumor histology.

Dr. Laury’s group used spatial transcriptomic analysis, which links variability in gene expression with heterogeneous tumor morphology, to evaluate the molecular features of microscopically indistinguishable WSI tissue regions previously identified by the AI model as highly associated with patient outcomes. 

The spatial transcriptomic profiles of these regions were evaluated and compared with the profiles of background tumor regions that the AI model did not select. Dr. Laury’s team validated these morphologic tumor region profiles and confirmed they are more predictive of patient outcomes than the background tumor regions. 

When these semi-opaque AI findings can be validated with molecular techniques, pathologists can learn to lean on them to guide clinical decision-making for treating high-grade serous carcinoma.

 

Implications of differential gene expression in outcome groups

Dr. Laury’s team’s analysis also identified underlying pathways that may drive the differential response to platinum-based chemotherapy in HGSC patients. Their findings demonstrate proof-of-concept for combining spatial transcriptomics and AI-based image analysis to investigate differential gene expression in HGSC.

“Focusing on specific tumor regions allowed us to pinpoint specific cell-cell interactions and upregulated genes that are not necessarily present throughout the tumor as a whole, but which may drive tumor progression,” Dr. Laury explains.

For instance, AI-detected regions in poor prognosis tumors express much higher levels of JUN, a proto-oncogene, and their cell subpopulations are characterized by a stress-associated transcriptomic signature. However, AI-detected regions in tumors with better outcomes contain tumor cell subpopulations comprising macrophages and plasma cell-depleted DNA repair signatures.

When comparing gene expression between patients with a short platinum-free interval (PFI), i.e., ≤6 months (PFI-S), and those whose PFI was ≥18 months (PFI-L), Dr. Laury’s team discovered that genes like JUN were differentially expressed. Here, JUN was exclusively upregulated in the morphologic regions detected by the AI model within the PFI-S tumors but not in the PFI-L. 

JUN is a known driver of carcinogenesis and therapy resistance and is a component of stress-related pathways, such as those enriched in the PFI-S morphologic regions. Dr. Laury’s group also confirmed these differential expression patterns by secondary assays like RNA in situ hybridization (ISH) staining, suggesting that JUN may be a clinical predictive marker in HGSC.

 

The future of AI-based histology as a prognostic tool

These findings demonstrate a synergy between spatial transcriptomics and AI-based histology image analysis and establish a more interpretable AI model to enhance future clinical translation. The proof-of-concept from Dr. Laury’s lab suggests there’s potential to develop clinical AI applications that help pathologists identify morphologic features otherwise indistinguishable by the human eye.

“This work supports the potential of incorporating AI-detection into the clinical setting and explores the possibility of using AI as a discovery tool to assist researchers combining morphological and molecular tumor features.” Dr. Laury


As an AI development tool, Aiforia® Create is versatile and easy to use, enabling pathologists to develop AI models to help them study complex biological systems like carcinomas, complement their existing suite of assays, and better understand how to guide patient care.

Learn more about what makes it unique from this blog post: Aiforia® Create: from research AI models to productization for clinical practice

 

References

Laury, A. R. et al. (2024, July). Opening the Black Box: Spatial Transcriptomics and the Relevance of Artificial Intelligence-Detected Prognostic Regions in High-Grade Serous Carcinoma. Modern Pathology, 37(7), 100508. doi.org/10.1016/j.modpat.2024.100508

Laury, A. R. et al. (2021, September 27). Artificial Intelligence-Based Image Analysis Can Predict Outcome in High-Grade Serous Carcinoma via Histology Alone. Scientific Reports 11, 19165. https://doi.org/10.1038/s41598-021-98480-0