A Multimodal Deep Learning Framework for Oral Cancer Prediction Using DenseNet-169 and Clinical Data Fusion

Abstract

Oral squamous cell carcinoma (OSCC) and leukoplakia with or without dysplasia require careful histopathological assessment. We examine a DenseNet-169-based multimodal framework combining H&E tissue images with clinical metadata for three-class oral lesion classification, emphasizing patient-independent validation rather than architectural novelty. The NDB-UFES cohort contains 237 paired images from 69 patients. We partitioned 55 patients (189 images) for development and 14 disjoint patients (48 images) for a locked internal test. Patient-grouped five-fold cross-validation supported image-only, clinical-only, fusion ablations, and frozen ImageNet backbone comparisons (DenseNet-169, ResNet-50, EfficientNet-B0). Internal validation Grad-CAM visualizations and clinical permutation sensitivity were exploratory. In controlled development ablations, the DenseNet-169 image-only branch attained mean patient-level accuracy 67.3% and macro-F1 0.612, versus 60.0% and 0.488 for fusion with a reducer. On the prespecified independent-patient internal test, the multimodal fusion model obtained patient-level accuracy 64.29% (9/14; exact 95% CI 35.14–87.24%), macro-F1 0.600 and macro one-vs-rest AUC 0.860; the image-only comparator obtained 57.14% (8/14), macro-F1 0.470 and AUC 0.756. Exact paired McNemar p=1.000 did not establish superiority. Fusion is feasible, but benefits remain uncertain owing to the small cohort, minority-class errors and absent external validation. The original submission's 87.18% image-level result arose from a patient-overlapping split and is not presented as an independent-patient estimate.