Hyperparameter optimized deep learning models for multiclass oral lesion classification in proprietary histopathological image Data

Oral cancer remains an important health issue in worldwide because of its significant incidence and mortality rates, with Squamous Cell Carcinoma accounting for more than 90% of incidents. Early detection of oral lesions is critical to prevent progression to advanced stages, yet traditional biopsy methods are invasive and prone to human error. To solve this, the paper introduces several key innovations like novel multi source dataset strategy combining large public and proprietary clinical data, comprehensive six-class classification covering the entire oral lesion progression spectrum and rigorous hyperparameter optimization with clinical validation. Deep learning framework uses cutting-edge CNNs like InceptionV3, ResNet50, DenseNet121, and MobileNet which are currently enhanced by increasing the data and hyperparameter modification. The model was tested on both public and private datasets. On the public dataset 4988 images: 2,494 normal, 2,494 OSCC, InceptionV3 performed the best with 96.2% accuracy and an AUC of 0.98. On the private dataset 2450 images: 1569 training, 489 testing and validation 392. DenseNet121 achieved 94.5% train accuracy and 90.6% validation accuracy, which shows strong generalizability. Our proposed method demonstrates a performance improvement of 4-6% in accuracy over greater robustness compared to other deep learning frameworks (91%), establishing new benchmark for automated OSCC classification. The model was optimized using the median filtering for reducing noise and augmentation techniques to increase its resilience. The framework presented here provides a non-invasive, an automated method of identifying OSCC early, minimizing diagnostic delay and leading to better patient results.