Dr Ayoub on an AI Model to Enhance Interpretability of Morphologic Heterogeneity in Glioblastoma

"What makes our algorithm different is that it's been trained in a supervised approach, where the ground-truth data set has been curated by [pathologists] with knowledge of the field, which makes the results more interpretable, as opposed to weakly supervised artificial intelligence models that would be sometimes less interpretable because of their less [curated training]."

Georges Ayoub, MD, MS, a postdoctoral research fellow at Dana-Farber Cancer Institute, discussed the development of a supervised deep-learning artificial intelligence (AI) model designed to evaluate morphological heterogeneity in glioblastoma. This approach contrasts with weakly supervised or unsupervised AI models, which have gained popularity for predicting molecular subtypes but often lack interpretability, he explained.

Traditionally, many AI tools in pathology have relied on unsupervised learning methods, where the algorithm makes predictions without labeled input from clinicians. Although efficient, these models have been limited by the absence of direct human oversight during training, potentially reducing their reliability in clinical decision-making, Ayoub stated. In contrast, supervised models use expert-labeled datasets—typically annotated by experienced pathologists—that provide more precise, reproducible, and interpretable results, he added.

Ayoub emphasized that the deep-learning algorithm developed by his team was trained using a robust, fully curated dataset created under the guidance of field experts. This ground-truth supervision allows for better alignment between algorithmic predictions and pathologist interpretations, offering increased transparency and clinical relevance. The supervised approach enables the algorithm to quantify spatial morphological heterogeneity—a hallmark of glioblastoma—and relate it to clinically significant outcomes.

The ability to distinguish tumor subregions based on histologic features has potential implications for both diagnostic accuracy and therapeutic stratification, Ayoub noted. By identifying areas of tumor with distinct morphological characteristics, the model can contribute to a more nuanced understanding of glioblastoma biology, which is known for its intratumoral variability and resistance to standard therapies.

As AI continues to evolve in oncology, supervised learning frameworks anchored in expert-defined annotations offer a path forward for developing interpretable, high-performance tools. These models may serve as critical adjuncts to traditional histopathology by enhancing the reproducibility and clinical utility of morphological assessments in glioblastom, he concluded.