Deep Learning-Based Multi-Class Cancer Image Classification: A Robust and Explainable Framework

Abstract

: Cancer is still one of the most common causes of death globally, and early and precise detection are crucial for enhancing the survival and treatment of cancer patients. The advent of deep learning has shown great promise for automation in cancer image analysis, but most models are poorly interpretable, widely fail to use global contextual information, and fail to produce the same results across different medical image datasets. The current research presents an Explainable and robust framework for multi-class cancer diagnosis called Deep Learning-Based Multi-Class Cancer Image Classification which integrates Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), feature fusion by attention mechanism, and Explainable Artificial Intelligence (XAI) techniques for accurate and interpretable cancer diagnosis. This proposed methodology starts from broad-based image preprocessing techniques such as image resizing, image normalization, image contrast enhancement, noise removal, and data augmentation, before conducting parallel feature detection using various CNN and Vision Transformer architectures. An adaptive attention module combines information from local and global feature representations to produce high-level discriminative features, and introducing visual and quantitative explanations for the classification decision, with Grad-CAM and SHAP improves interpretability and trustliness for the medical reader. The proposed framework was tested on publicly available image datasets related to cancer including BreakHis, LC25000, ISIC, PatchCamelyon and TCGA with the same set of experimental conditions, while compared with state-of-the-art deep learning models: VGG16, ResNet50, DenseNet121, EfficientNet-B3, and Vision Transformer. Various performance metrics like Accuracy, Precision, Recall, F1 score, Specificity, ROC AUC and Confusion Matrix analysis were used in the experimental evaluation. The performance of the proposed framework was outstanding with an overall classification accuracy of 98.46%, precision of 98.29%, recall of 98.11%, F1-score of 98.20% and a ROC-AUC of 0.997 which was better than the already existing benchmark machine learning models. The effectiveness of the attention-based feature fusion module was validated by ablation studies, which showed that the combinations of the two architectures CNN and Vision Transformer to be able to improve the feature representation. Moreover, clinically meaningful visual explanations are generated by the Explainable AI, but with no loss of predictive performance. The obtained results show that the proposed approach is powerful, scalable and is very suitable for future clinical cancer diagnosis support systems as well as intelligent computer-aided diagnosis.