Authors: Harshitha T N, Soujanya R
Abstract: Oral cancer is a critical and life-threatening disease, where early detection plays a vital role in improving patient survival rates. However, traditional diagnostic approaches rely heavily on manual clinical examination and biopsy, which are time-consuming, invasive, and often lead to delayed diagnosis. To address these limitations, this paper proposes a deep learning framework for automated oral cancer detection using medical image analysis and lesion-focused classification techniques. The proposed system integrates image preprocessing, lesion segmentation, and deep convolutional neural networks (CNNs) for accurate classification. Preprocessing techniques such as contrast enhancement and noise reduction are applied to improve image quality. Lesion regions are extracted using Otsu thresholding and contour-based segmentation to isolate regions of interest (ROI), which enhances feature learning. Multiple deep learning architectures, including Baseline CNN and EfficientNet-B0 are evaluated for performance comparison. In addition, the proposed framework integrates lesion segmen-tation and deep feature extraction to improve classification robustness and diagnostic performance. To enhance model interpretability, Grad-CAM is employed to visualize the regions contributing to predictions, making the system more transpar-ent for medical applications. Experimental results demonstrate that the proposed EfficientNet-B0 based model achieves superior performance compared to baseline approaches, with improved accuracy and F1-score on the test dataset. The proposed framework provides an efficient, scalable, and interpretable solution for early-stage oral cancer detection, supporting clinical decision-making and reducing diagnostic delays.
