AI-based classification of laryngeal dysplasia and lymphocytic activity quantification from routine histology.

Abstract

6080 Background: Laryngeal Dysplasia (LD) is a premalignant condition arising in the lining of the larynx. It is graded based on cytological and architectural features present in the epithelium of H&E-stained histology images. However, LD grading suffers from high inter- and intra-rater variability and is not always predictive of malignant transformation. Additionally, distinguishing LD from other laryngeal lesions, such as squamous cell carcinoma (SCC) or benign polyps, remains challenging. Artificial intelligence (AI) offers a solution by enabling objective classification of lesion types, whilst identifying key features associated with LD progression, including lymphocytic infiltration. We propose an AI model using weakly-supervised deep learning to classify LD and highlight potential diagnostic features. Methods: We used 109 H&E-stained whole slide images (WSIs) from 82 cases (UHCW and Dundee) scanned at 40× magnification (0.12 microns-per-pixel, mpp) using a Pannoramic 250 whole-slide scanner. The dataset comprised 50 LD cases (65 WSIs), 20 laryngeal SCC cases (28 WSIs), and 12 benign polyp cases (16 WSIs). Using a pre-trained H-optimus-0 model, we extracted patch-level (224×224 pixels) features from the slides (20× magnification, 0.5 mpp), with a TransMIL aggregator predicting slide-level classifications for dysplasia, SCC, and polyps. Additionally, we derived slide-level intra-epithelial lymphocyte (IEL) and peri-epithelial lymphocyte (PEL) scores using HoVer-NeXt based lymphocyte segmentation, in and around the epithelium (segmented by HoVer-Net+) in LD cases, and compared these scores across WHO LD grades using Mann-Whitney U tests. Results: In Monte Carlo cross-validation experiments (10 repeats), the model achieved an average one-versus-all AUROC of 0.85 and AUPRC of 0.73 for lesion classification (dysplasia vs SCC vs polyp). In LD cases, both IEL and PEL scores were significantly higher in severe dysplasia cases compared to moderate (IEL: r rb = 0.09, p = 0.02; PEL: r rb = 0.36, p = 0.02) and mild dysplasia (IEL: r rb = 0.09, p = 0.01; PEL: r rb = 0.36, p = 0.003). This suggests a potential link between increased lymphocyte presence (activity) and higher grades of dysplasia. Conclusions: We present a novel AI model for classifying laryngeal lesions and quantifying lymphocytic activity in LD. Our findings suggest the diagnostic potential of AI in identifying LD, whilst highlighting peri- and intra-epithelial lymphocyte density as a potential biomarker, which has not been previously linked to dysplasia grade. Further validation in large, multi-centric datasets is required.