MedFoundX: a foundation model for biomedical image classification and segmentation

Abstract

The demand for efficient and generalizable artificial intelligence models in medical imaging is rapidly increasing. This thesis introduces MedFoundX, a unified foundation model designed for classification and segmentation across various biomedical imaging modalities. The model’s architecture features a pre-trained EfficientNet-B3 backbone, integrated with Convolutional Block Attention Modules (CBAM) and Multi-Head Attention. A sequential weight transfer training protocol is applied to eight publicly available and clinically relevant datasets, encompassing multiple imaging types, including MRI, CT, X-ray, and colonoscopy. MedFoundX achieved nearly perfect classification performance in several datasets, significantly exceeding established models such as CNN, KAN, ResNet-50, Swin-Transformer and VGG-16. In segmentation tasks, it reached mean Dice coefficients of up to 0.964 on the MedSeg-Liver dataset and 0.941 on the Kvasir-SEG dataset, considerably outperforming other models like CNN, KAN, and DeepLabV3. Furthermore, MedFoundX was tested on two unseen datasets with fine-tuning. It achieved an impressive accuracy of 98.5% on the unseen PMRAM classification dataset, with only six misclassifications out of 400 scans. In the CVC-ClinicDB segmentation dataset, MedFoundX recorded a mean Dice score of 0.83 during inference, along with high sensitivity (0.94) and specificity (0.988). Computational investigation revealed that MedFoundX has 22.21 million parameters, requires 2.8 billion FLOPs, and occupies 84.7 MB of memory, allowing real-time inference and outperforming larger models such as ResNet-50, DeepLabV3, and VGG-16. This work effectively addresses the performance-efficiency gap in clinical AI by providing a single, attention-enhanced architecture that generalizes well across various modalities and tasks, achieving state-of-the-art classification and segmentation without compromising computational feasibility. Its successful application to unseen datasets demonstrates its robust generalization capabilities.