Leveraging Pretrained Models for Multimodal Medical Image Interpretation: An Exhaustive Experimental Analysis

Artificial intelligence (AI) in radiology, particularly pretrained machine learning models, holds promise for overcoming image interpretation complexities and improving diagnostic accuracy. Although extensive research highlights their potential, challenges remain in adapting these models for generalizability across diverse medical image modalities, such as Magnetic Resonance Imaging (MRI), Computed Tomography (CT), and X-rays. Most importantly, limited generalizability across image modalities hinders their real-world application in diverse medical settings. This study addresses this gap by investigating the effectiveness of pretrained models in interpreting diverse medical images. We evaluated ten state-of-the-art convolutional neural network (CNN) models, including ConvNeXtBase, EfficientNetB7, VGG architectures (VGG16, VGG19), and InceptionResNetV2, for their ability to classify multimodal medical images from brain MRI, kidney CT, and chest X-ray (CXR) scans. Our evaluation reveals VGG16s superior generalizability across diverse modalities, achieving accuracies of 96% for brain MRI, 100% for kidney CT, and 95% for CXR. Conversely, EfficientNetB7 excelled in brain MRI with 96% accuracy but showed limited generalizability to kidney CT (56% accuracy) and CXR (33% accuracy), suggesting its potential specialization for MRI tasks. Future research should enhance the generalizability of pretrained models across diverse medical image modalities. This includes exploring hybrid models, advanced training techniques, and utilizing larger, more diverse datasets. Integrating multimodal information, such as combining imaging data with patient history, can further improve diagnostic accuracy. These efforts are crucial for deploying robust AI systems in real-world medical settings, ultimately improving patient outcomes.