What question did this study set out to answer?

This research aims to develop a non-invasive diagnostic tool for early detection of oral squamous cell carcinoma (OSCC) using AI and histopathological imaging.

January 17, 2026Open Access

AI-Assisted Diagnosis of Oral Cancer: A Deep Learning Approach for Early Detection Using Histopathological Imaging

Puntos clave

This research aims to develop a non-invasive diagnostic tool for early detection of oral squamous cell carcinoma (OSCC) using AI and histopathological imaging.
Developed a deep learning model based on transfer learning from pre-trained convolutional neural networks (CNNs).
Utilized image preprocessing techniques like resizing, normalization, and brightness adjustments to enhance dataset features.
Implemented a soft voting ensemble method to boost model accuracy and reduce false negatives.
Assessed model performance using metrics like accuracy, F1-score, recall, precision, and ROC-AUC.
Applied Grad-CAM to visualize critical lesion regions for better interpretability.
The proposed model provides a cost-effective and non-invasive tool for early OSCC detection.
Achieved improved accuracy and robustness through ensemble methods and image enhancements.
Enabled more reliable risk assessments with enhanced clinical trustworthiness from interpretability techniques.

Resumen

The prevalence of OSCC (oral squamous cell carcinoma) continues to be a leading cause of death from cancer world-wide. Even though substantial developments have occurred in medicine, OSCC can often be diagnosed too late. In addition, due to the paucity of screening facilities are available, especially in under-developed areas, Through the utilization of invasive diagnostic procedures such as biopsies continue to be the primary means of diagnosing OSCC. As such, the need for developing non-invasive diagnostic tools continues to grow. In this paper, we present an artificial intelligence-based deep learning framework that uses histopathological images to detect and classify OSCC. As such, the imperative of developing non-invasive diagnostic tools continues to grow. This model is founded upon the principle of transfer learning from the pre-trained CNNs (Convolutional Neural Networks) MobileNet, VGG16, and ResNet5, and were trained on curated datasets of histopathological images of oral tissues classified into normal, pre-cancerous and cancerous. We also employed numerous image preprocessing techniques (resizing, normalization, rotation, flipping, and brightness adjustments) to augment the features of our dataset, which was necessary to strengthen the learning The diagnostic utility of the proposed model and to increase its feature robustness. To further improve the accuracy and reduce false negative results, we implemented a soft voting ensemble method. Soft voting is a weighted version of hard voting and combines the probabilistic output values of all of the separate individual CNN models. Commonly adopted performance metrics for classification models, namely accuracy,F1-score,recall, precision and ROC-AUC were implemented to assess the model’s overall performance. Additionally, Grad- CAM (Gradient-weighted Class Activation Mapping) was leveraged to visually demonstrate the most critical lesion regions with an intention to enhance the interpretability and clinical trustworthiness of the proposed model. The model was developed using both TensorFlow and PyTorch on a Google Colab GPU environment and presents a non-invasive, cost-effective, and explainable diagnostic tool to assist clinicians in performing early oral cancer screening and risk assessments by integrating AI-driven image analysis with clinical oncology. This work provides a foundation for future multimodal and transformer-based diagnostic systems, capable of transforming the early detection and handling of oral cancer cases.

Leer artículo completoexternamente

Preguntar a la IA

Me gusta

Guardar

Ver artículo completo