Leveraging AI and Machine Learning for Early Detection of Oral Squamous Cell Carcinoma with EfficientNetB3

Authors: Raja Muppidi, Bhuvana Murki

Abstract

This project delves into the effectiveness of the EfficientNetB3 model in detecting Oral Squamous Cell Carcinoma (OSCC) from histopathological images. The study evaluates the model’s adaptability and performance metrics, including training and validation accuracies over several epochs, emphasizing its potential as a crucial tool in early cancer detection. Our project aims to bridge the gap between traditional diagnostic methods and advanced AI-driven techniques, ensuring faster and more accurate diagnosis.

Introduction

Oral Squamous Cell Carcinoma (OSCC) is a cancer affecting the mouth and surrounding tissues. Early detection is crucial for effective treatment, but it often relies on conventional methods that can be slow and subjective. This project introduces an AI-based model, EfficientNetB3, tailored to enhance the diagnostic process by providing rapid and accurate results through image analysis. The EfficientNetB3 model detected Oral Squamous Cell Carcinoma (OSCC) from histopathological images. This approach has shown promising results in detecting OSCC.

Methodology

The methodology focused on leveraging the EfficientNetB3 architecture to process histopathological images effectively. The process included:

• Model Configuration: An EfficientNetB3 backbone was employed and fine-tuned for our specific task without the top layers, allowing custom layers to be added for classification.
• Data Handling: Training involved a dataset of labeled images (OSCC and normal tissues). Techniques like data augmentation (rotation, zoom, and flipping) were used to enhance the model’s ability to generalize.
• Model Compilation and Training: The model was compiled with the Adamax optimizer and categorical cross-entropy loss function, focusing on optimizing accuracy. The training was conducted over 100 epochs with early stopping to prevent overfitting

Results

The training phase showed progressive improvement in model accuracy and a decrease in loss, both in training and validation sets:

• Initial Training: The model initially struggled, with accuracy hovering around 53-54% in early epochs.
• Progressive Adaptation: Substantial improvements were evident as the model adapted, with accuracy eventually reaching 86.69% by the 83rd epoch.
• Final Metrics: The best validation accuracy achieved was 96.53%, with the corresponding loss significantly reduced, highlighting the model’s capability to efficiently learn from the data.

Discussion:

The model’s training and validation curves indicate a consistent improvement, showcasing the efficiency of the EfficientNetB3 architecture in learning complex patterns within histopathological images. The early epochs indicated a struggle to adapt to the nuances of the data, but through continuous learning and model adjustments, it achieved high accuracy levels. The detailed performance metrics are as follows

• Highest Validation Accuracy: 96.53% in the later epochs.
• Lowest Validation Loss: Achieved concurrently with the highest accuracy, suggesting the model’s robustness.

These results underscore the potential of integrating EfficientNetB3 into clinical workflows to assist pathologists, thereby reducing diagnosis time and increasing the accuracy of OSCC detection.

Conclusion:

Our Project demonstrates the potential of using EfficientNetB3 in the medical field as a reliable tool for OSCC detection. The success of this model could pave the way for broader applications of AI in histopathological analysis, making it a cornerstone for future diagnostic processes.

Future Work:

Future research will focus on refining the model through real-world clinical trials and expanding its application to other types of cancers. Further advancements might also explore real-time processing capabilities to provide immediate diagnostic support in clinical settings.