Automated detection of basement-membrane breaches in oral squamous cell carcinoma : A comparative study of convolutional neural network architectures

Background: Oral squamous cell carcinoma (OSCC) remains a leading cause of cancer-related morbidity and mortality worldwide. The ability to detect early histopathological evidence of invasion-particularly basement membrane (BM) discontinuity at the epithelium-connective tissue interface-is crucial for timely intervention; yet, this subtle feature is often overlooked in automated image analysis research. This study aimed to develop and benchmark convolutional neural network (CNN) models for the automated detection of BM breaches in OSCC, introducing the first publicly available, expert-annotated dataset dedicated to this diagnostic hallmark. Methods: We assembled 66 high-resolution hematoxylin-eosin (H&E) stained histology images, evenly split into "break" (BM discontinuity with tumour invasion) and "normal" categories. Expert oral pathologists annotated the epithelial-connective tissue interface, producing region-of-interest metadata for each image. The dataset was divided into training, validation, and test sets. Five CNN architectures were evaluated under same training conditions: two custom models (SimpleCNN and DeepCNN) and three transfer learning models (ResNet-18, MobileNetV2, EfficientNet-B0). Performance was measured by test accuracy and AUC. Results: ResNet-18 was best performing CNN with 86.67% accuracy and 0.87 AUC, effectively discriminate the lesion with break from normal mucosa. MobileNetV2 followed with 80% accuracy (AUC 0.83). SimpleCNN, DeepCNN, and EfficientNet-B0 achieved about 73%. Overall, transfer learning models outperformed custom architectures, highlighting the benefits of pre-trained features in this histopathological task. Conclusions: This study highlights AI's role in detecting epithelial-basement membrane breaks and a key microscopic feature for diagnosing invasion, demonstrating how deep learning enables precise, consistent, objective evaluations.