XAI BASED OSCC DETECTION SYSTEM

N.B:- In anticipation of publishing our paper, we have maintained confidentiality of our codes and in depth detailings until the time of publication.

Problem Statement

Using histopathological images, this research aims to develop an explainable AI-driven model that combines Convolutional Neural Networks (CNNs) with interpretable techniques. The goal is to enable accurate and transparent classification of oral squamous cell carcinoma (OSCC), providing clinicians with interpretable insights into the model’s decision-making process and enhancing trust, confidence, and clinical acceptance in OSCC diagnosis.

Objective

The question is, can we really rely on an artificial intelligence system which will give a yes or no answer, without knowing why or how the answer is being given? We believe the answer is a pretty strict no. Because, one the first hand, many of the normal people aren’t really aware clearly, what an AI system is, and secondly, even many technical people don’t have a clear idea about how an answer is being generated in a model. This is where the major motivation of the work came from. If a model can give an answer about whether anyone has cancer or not, it also should be able to say “why” such an answer is being given. We aim to do exactly this with our built model.

Images from the OSCC Dataset

Here, we tried to train out CNN models with a dataset that contains histopathological imaages. After the training session completed, the last layer of the CNN model was implemented with XAI models. Explainable AI generates a visual system that clearly indicates the reasoning in favor of the given result.

Dataset

For grading cancer, histopathological image analysis is frequently performed all around the world. Histopathology slides offer more detailed diagnosis information than mammography, CT, and other imaging tests. It is among the cheapest morphological techniques. Rapidly and with little risk to the patient, samples can be obtained to make the images. As a good dataset is significant for the practical outcome of any model, in fields like cancer detection, histopathological images have proved their competency before. Due to that, we used a publicly available dataset to detect cancer in the oral squamous cell. The dataset is called "Oral cancer histopathological dataset".

Images from the OSCC Dataset

To enrich the training data, increase model generalization, improve robustness, and mitigate some issues, we performed a few augmentation techniques on the stratified samples.

Samples from the augmented dataset

Methodology & Experimentations

Training approaches used: -

• Fine-tuning
• Cost-Sensitive Approach
• Contrastive Learning Approach
• Triplet contrastive Loss
• Max-Margin contrastive Loss
• Supervised contrastive Loss

CNN based models used to train: -

• AlexNet
• DenseNet-121
• InceptionV3
• MobileNetV2
• ResNet-50
• VGG-16
• VGG-19

Architecture of the fine-tuned approach

During the training process, we employed a strategy of freezing and unfreezing certain layers to achieve better results. This approach allowed us to selectively update the weights of specific layers while keeping others frozen. By controlling which layers were trainable, we aimed to strike a balance between leveraging pre-trained knowledge and allowing the model to adapt and learn new representations for the specific task at hand. Subsequently, various XAI (Explainable Artificial Intelligence) techniques were applied to assess the interpretability of the models.

Model Evaluation

The performance of the fine-tuned models on individual class along with accuracy

For our experiments, we utilized fine-tuned settings for all models except for AlexNet, which was trained from scratch. The training process for each model required multiple epochs to reach convergence. Among the models, VGG-16 and DenseNet-121 demonstrated the most favorable results, while ResNet-50’s performance was unsatisfactory. AlexNet did not exhibit the best performance either, while the results for the remaining models fell somewhere in between. Given the imbalanced nature of our dataset, we observed a noticeable difference between the precision and recall values for both classes in all models, except for VGG-16.

Model Interpretation

XAI techniques used: -

• Gradient based methods:
  • Grad-CAM
  • Grad-CAM++
• Gradient free methods:
  • Score-CAM
  • Faster Score-CAM
• Perturbation based method: LIME

XAI methods applied on VGG-16 for a correctly classified ’OSCC’ image

For the fine-tuned versions of VGG-16 and its cost-sensitive variations, all XAI methods identified the keratin pearl as a crucial factor in determining the "OSCC" class. In the case of contrastive learning, the version trained with max-margin loss yielded interpretability with grad-cam and faster-scorecam methods. The scorecam and its updated version produced consistent pixel contributions for the triplet loss. In the supervised contrastive loss scenario, every pixel seemed to play a role in declaring the class.

LIME applied on all models for a correctly classified ’OSCC’ image

Different models and their versions were evaluated for their ability to classify images as ’OSCC’. The evaluation involved analyzing the contribution of positive and negative superpixels in determining the class.

Conclusion

We have explored three distinct approaches in our models: fine-tuning, cost-sensitive learn�ing, and contrastive learning. Each technique yielded varying outcomes, with some models exhibiting improved performance through fine-tuning, while others benefited from class weight adjustments. Similarly, different contrastive loss functions resulted in different levels of performance. Through the application of Explainable AI, we have generated visualizations that highlight the models’ behavior. These visualizations have revealed that certain models excel at identifying smaller portions, while others struggle to locate crucial areas. Notably, the perturbation-based technique known as LIME has demonstrated superior visu�alization capabilities compared to other methods.