An Empirical Analysis of Transformer-Based and Convolutional Neural Network Approaches for Early Detection and Diagnosis of Cancer Using Multimodal Imaging and Genomic Data

Early diagnosis of cancer has focused on the use of advanced algorithms to achieve accurate diagnosis. The proposed study assesses the effectiveness of Transformer-based models and Convolutional Neural Networks (CNN) in cancer diagnosis with respect to multimodal imaging and genomic data. The performance comparisons between the two algorithmic methods with such complex datasets, which combine multi-modal imaging and genomic information, are presented. In search of the optimal neural network configuration, a series of experiments were conducted with respect to different layers, attention mechanisms in case of transformers, and convolutional architectures in case of CNNs. Besides, parameters related to training, such as learning rates, batch sizes, and optimization algorithms, have also been systematically tuned. The different models were evaluated against accuracy, precision, recall, and the F1-score. Our results show that the proposed multimodal model, with accuracy from 92.5 to 93.2, F1-scores between 91.5 and 92.2, precision of 91.5 to 92.2, and recall values of 92.5 to 93.2. In contrast, much lower accuracy, F1-scores, precision, and recall values were noticed when using baselines, especially VGG. All these findings indicate the fact that the presented techniques, especially the Multimodal and Transformer models, are more robust solutions for classification tasks with better balance between precision and recall, as well as with higher overall accuracy. This came with the cost of the expense of computational resources: CNNs are less resource-intensive but have competitive performance with better precision and recall. The results underline how algorithm selection and hyperparameter optimization play a crucial role in cancer detection tasks. This study has shown how state-of-the-art deep learning methods can be effectively combined with multi-modal data for building more accurate and efficient systems in cancer diagnosis. Two main lines of future work would be improving these algorithms and understanding their applicability in real clinical practice to obtain maximum benefits from them.