Article’s

Medical image classification is a critical component in computer-aided diagnosis, yet existing deep learning models often struggle with generalization across diverse imaging modalities. This paper proposes a Hybrid Convolutional Neural Network (HybridCNN) integrated with a Global Token Mixing mechanism to address both local and global feature extraction challenges. The model is trained on the full MedMNIST dataset, enabling exposure to a wide range of medical imaging types including X-rays, histopathology images, and ultrasounds. The architecture combines convolutional layers for fine-grained local feature extraction with token-based global interaction layers inspired by Vision Transformers and MLP-Mixer models. Experimental evaluation demonstrates that the proposed approach improves contextual understanding, enhances classification robustness, and supports scalable medical image analysis across heterogeneous datasets. The system is further validated through an inference pipeline capable of real-time predictions with confidence estimation.