Designing a Hybrid CNN-Transformer Architecture for Robust Medical Image Diagnosis

Medical image diagnosis increasingly relies on deep learning against a background where convolutional neural networks capture local texture well but miss long-range spatial relationships, while transformer models need large labeled datasets and heavy computation that clinical settings rarely have. This study pursues the objective of designing a hybrid CNN-Transformer architecture that fuses convolutional feature extraction with self-attention to raise diagnostic accuracy on X-rays, MRI scans, and histopathology slides without exceeding the computational budget of real-world deployment. The proposed method pairs a lightweight convolutional backbone, which extracts local texture and edge features, with a transformer encoder that models global relationships across image regions, merging the two branches through a cross-attention fusion module before a classification head outputs the diagnosis. Training the architecture on publicly available medical imaging datasets and evaluating it against standalone CNN and transformer baselines using accuracy, sensitivity, specificity, and F1-score, the result reveals that the hybrid architecture beats both baselines, raising accuracy and sensitivity on cases with subtle abnormalities that need local texture recognition and broader spatial reasoning together, while inference time stays within a practical range for clinical workflows. These findings lead to the conclusion that the hybrid CNN-Transformer design gives clinicians a computationally feasible path to more robust medical image diagnosis, and the paper releases the architecture and source code to support reproducibility and further research toward accessible, trustworthy AI-assisted healthcare.