AI in Healthcare Imaging

Abstract

AI-based medical image analysis has moved from research curiosity to clinical deployment in radiology, pathology, ophthalmology, and dermatology at remarkable speed. Yet clinical adoption has exposed gaps between benchmark performance and real-world utility: models trained on curated research datasets degrade under acquisition protocol variation, exhibit demographic performance disparities, and lack the calibrated uncertainty outputs that clinicians need to decide when to trust AI recommendations. This study evaluates nine deep learning architectures for medical image analysis across five imaging modalities: chest X-ray (pneumonia, effusion, pneumothorax detection), brain MRI (tumour segmentation), retinal fundus (diabetic retinopathy grading), dermatoscopy (melanoma vs. benign lesion), and histopathology whole-slide images (colorectal cancer grading). Architectures evaluated include ResNet-50, EfficientNet-B7, Vision Transformer (ViT-L/16), Swin Transformer, Medical SAM (Segment Anything Model adapted for medical imaging), nnU-Net, DenseNet-121, BioViL-T (vision-language model), and an ensemble. Datasets span 14 institutions across six European countries (n=284,640 images). ViT-L/16 and Swin Transformer achieve AUC >= 0.94 on chest X-ray pathology detection across most findings. Medical SAM achieves the highest Dice coefficient on brain tumour segmentation (0.924). BioViL-T enables zero-shot detection for rare findings absent from training data. Demographic performance gaps (AUC difference by sex: up to 0.048; by age group: up to 0.064) are identified across all models. A clinical deployment readiness framework covering performance validation, bias auditing, and uncertainty calibration is proposed.