Harnessing deep learning for enhanced detection of thymic epithelial tumors

The increasing reliance on deep learning for medical image classification has significantly improved tumor detection. However, ‎traditional deep learning models often suffer from redundant feature extraction, poor feature separability, and suboptimal classifica-tion ‎performance. This study proposes an advanced hybrid framework integrating Adaptive Neuro-Evolutionary ‎Pruned Embedding (ANPE) and Self-Supervised Embedding Transformation (SSET) for tumor classification to overcome these limitations. The ANPE module ‎optimizes feature selection by employing multi-scale feature extraction, neuro-evolutionary pruning, and self-attention ranking to retain only ‎the most discriminative features, thus enhancing classification accuracy. The SSET module, on the other hand, leverages contrastive learning ‎techniques, specifically SimCLR (Simple Contrastive Learning) and MoCo (Momentum Con-trastive Learning), to improve feature ‎representation and cluster separation. The proposed ANPE-SSET framework was evaluated on the TCGA-THYM dataset, a benchmark ‎dataset for histopathological tumor classification. Experimental results demonstrate that ANPE-SSET outperforms baseline models such as ‎CNN, ResNet, and Vision Transformers (ViTs). The proposed model achieved an accuracy of 96.2%, significantly surpassing CNN ‎‎(85.3%), ResNet (90.5%), and ViT (92.1%). The AUC-ROC score of 98.3% high-lights its superior ability to distinguish between tumor ‎classes. In conclusion, the ANPE-SSET hybrid model effectively integrates evolutionary feature selection and self-supervised contrastive ‎learning, leading to state-of-the-art performance in tumor classifica-tion. The results demonstrate its superiority over traditional models, ‎making it a promising approach for medical image analysis. Future work will focus on extending this framework to other medical imaging ‎datasets and further optimizing it for clinical applica-tions‎.