A Fine-Tuned YOLOv5 and Exception Model for Oral Cancer Detection

Oral cancer remains a significant concern in public health, and early diagnosis is vital for adequate intervention. In this study, we present a dual-model approach of oral cancer detection via two methods—the Xception model and the YOLOv5 model. This dataset has pictures on oral lesions categorized into malignant and benign. Xception model is known for efficient feature representation, and this oral cancer detection shows promising results with 96.73% accuracy of this model at epoch 30 showing sufficient power to differentiate between malignant vs benign cases. Getting closer to an impressive accuracy of 97.84% at epoch preferably by the Xception model as training. Similarly, YOLOv5 is a powerful object detection model that harnesses its power to detect and classify lesions in images. YOLOv5 model gets an accuracy of 76.54% on epoch 38. It is to be noted that as you train YOLOv5, it shows significant improvement and by epoch 38 reaches an accuracy of 85.44
^%
. Comparative study reveals that the Xception model is more accurate than YOLOv5 across all stages of training and testing epochs. The Xception model outperforms consistently and continually, indicating the potential to diagnose between malignant or benign oral lesions. These outcomes highlight the need to use appropriate models for specific medical imaging tasks, and it is shown that Xception serves as a reliable model candidate when diagnosing oral cancer. This study offers important information regarding utilization of deep learning models for oral cancer detection, revealing minor distinctions in effectiveness between the Xception and YOLOv5 architectures. The yielded results in terms of the accuracies achieved at different epochs provide a detailed understanding about how these models learn and thus whether they can help clinicians diagnose oral cancer early enough and with great accuracy.