From a pathologist’s point of view, H&E (haematoxylin & eosin) staining is an essential tool in data analysis but how does it work in data-rich imaging? John Le Quesne, Professor in Molecular Pathology at the University of Glasgow takes us through the typical journey of a pathologist. Firstly, one examines a tray of slides and ponders what they are looking at and mentally integrates and observes the large number of variables in the images. The information is passed through various checkpoints and ultimately boils down to 10 words that have an impact on patient management.
From Le Quesne’s perspective, the main frustration lies with the fact that all information left on the slide with potential value, may not be available for subsequent use clinically or in a research setting. Furthermore, the image data is so complex that it may be beyond human interpretation. To tackle this, Le Quesne introduced a new self-supervised AI method called HPL (histomorphological phenotype learning), which discovers morphological features in histopathology images without supervision.
Moreover, the SHAP analysis showed that the HPL was able to identify clusters associated with good or bad outcomes. Poor outcome clusters were characterized by poorly differentiated solid growth patterns and immunologically cold environments, while good outcome clusters were low-grade and immunologically hot. Upon further testing, the unsupervised model outperformed the gold standard supervised model in prognosticating lung adenocarcinoma outcomes, achieving a C index of 0.73 compared to 0.68 for the supervised approach.
Le Quesne pointed out the significance of this technology: “What's surprising is that we're being shown these morphologies and they're being linked to a clinical ground truth with really no human expertise. So, in an afternoon, it's showing us the survival associations of morphologies which have taken us over 100 years to discover painstakingly by hand.”
Next, Le Quesne wanted to investigate the molecular implications and meaning behind these clusters. The method was successfully applied to a richer data set of 7-channel images, comprising T cell subtypes, macrophages, SMA, and T67. It demonstrated impressive prognostication from single one-millimetre cores of tissue per patient. The method was applied to malignant mesothelioma and demonstrates impressive disease subtyping and prognostic value.
In terms of future direction, the HPL displays vast biomarker potential and can stratify patients using clinically manageable low Plex Multiplex chromogenic methods alongside H&E. The team is working on incorporating large language models and high-resolution images to build on their method.
