An AI-based platform using discrete-frequency infrared microscopy detected biomarker-associated molecular signatures in unstained tissue from oral potentially malignant dysplasia, identifying cases at higher risk of becoming malignant.  

The approach enables simultaneous morphologic assessment and biomarker profiling of oral potentially malignant dysplasia from unstained tissue, with the potential to improve prognostication and prediction of malignant transformation, according to findings presented in a late-breaking abstract at the American Head & Neck Society (AHNS) 12th International Conference on Head and Neck Cancer.  

Background

Oral potentially malignant disorders are the precursors of more than 80% of oral cancers.  

Moni Abraham Kuriakose, MD, FDSRCS, FFDRCS, FRCS Ed, FRCS, BDS, Director of Translational Research for Head & Neck/Plastic & Reconstructive Surgery, Director of the Skull Base Program, and Vice Chair of the Department of Head & Neck/Plastic & Reconstructive Surgery at Roswell Park Comprehensive Cancer Center, explained that predicting which patients with dysplasia will progress or transform to oral cancer remains challenging because disease progression is highly variable and only weakly correlated with dysplasia grade. Hematoxylin and eosin (H&E) histopathology is the current standard for identifying precancerous lesions, but it is associated with substantial interobserver variability and does not capture the underlying molecular changes that drive malignant transformation.  

Although the prognostic biomarkers S100A7 and podoplanin and the diagnostic biomarkers CD44 and SNA-1 are associated with a higher risk of malignant transformation or poor survival, they cannot accurately predict disease transformation.  

Discrete-frequency infrared imaging identifies the chemical composition of tissue based on its unique infrared spectral signatures. “The problem was we did not have a mechanism to analyze the spectra—the massive data coming out of it, so we have used AI to characterize solutions,” Dr. Kuriakose told ASCO AI in Oncology.  

By incorporating AI workflows, the imaging platform can map biomarker-associated molecular signatures of oral potentially malignant dysplasia directly from unstained tissue.  

The benefit of this approach, Dr. Kuriakose explained, “is that it is very, very cost-effective. We don’t have to extract DNA or RNA or perform sequencing.”  

Study and Model Methods 

The researchers used H&E-stained tissue, biomarker-specific immunohistochemistry, and corresponding ground-truth biomarker labels to train the AI model. Annotation maps served as training labels for the classification and segmentation models.  

They developed a dual-branch neural network that integrates darkfield morphology and infrared imaging for multilabel biomarker classification. The multimodal deep learning model included automated tissue segmentation, patch extraction, patch labeling, and classification. The model generated expression scores for each biomarker from the infrared images.  

Feature selection was optimized using signal-to-noise ratio screening and further refined with Squeeze-and-Excitation modules and a knockoff procedure to minimize false discoveries.  

The testing cohort included 32 patients: eight with severe dysplasia, eight with moderate dysplasia, six with mild dysplasia, five control individuals, and five patients with oral squamous cell carcinoma.  

Study Findings 

The AI-assisted microscopy approach achieved an overall accuracy of 94.5% and an F1 performance score of 0.823 (± 0.019).  

Infrared spectral signatures were independent predictors of disease progression.  

Biomarker classification performance was robust across all four biomarkers. The model achieved an area under the curve of 0.853 for podoplanin, 0.912 for S100A7, 0.843 for SNA-1, and 0.875 for CD44. It also mapped the spectral features of each biomarker, including their localization and intensity.  

The AI-assisted platform enabled the simultaneous assessment of morphology and biomarker status from a single unstained tissue section.  

The researchers said the approach could improve risk stratification and prediction of malignant transformation in oral potentially malignant dysplasia.  

Going forward, Dr. Kuriakose said the research team plans to improve the model’s diagnostic accuracy by integrating clinicopathologic variables and expanding the testing cohort with additional patients from Roswell Park’s database.  The researchers also aim to reduce the time required for infrared imaging and data acquisition.  

He added that as infrared spectroscopy is increasingly being investigated in other cancer types, including prostate, breast, and colorectal cancers, the platform may have broader applications.  

The team also hopes to develop a unified decision support algorithm that can be readily adopted in clinical practice.  

DISCLOSURES: The study authors reported no disclosures. 

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.