An Analysis of Multi-Criteria Performance in Deep Learning-Based Medical Image Classification: A comprehensive review

Objective: Investigate the potential of deep learning, specifically Convolutional Neural Networks (CNNs), for disease detection and classification in medical image analysis. Methods and procedures: Review various CNN architectures (ResNet, EfficientNet, MobileNet, VGG, DenseNet) on retinal fundus and brain tumor datasets, evaluating performance through loss, AUROC, accuracy, and precision. Results: Retinal fundus: ResNet 152 achieved highest accuracy and precision, but also highest loss. Precisions indicate potential for improvement. Brain tumor: VGG 16 achieved highest accuracy and precision, but EfficientNet models displayed poor differentiation between positive and negative classes. CheXpert: MobileNet demonstrated best overall performance. No single model consistently outperformed others across all datasets. Conclusion: Deep learning shows promise for medical image analysis, with varied performances across different architectures and datasets. No single ”best” model exists, necessitating careful selection based on specific disease and data characteristics. Precision needs improvement in many cases, highlighting an area for further research. EfficientNet models require further investigation for brain tumor analysis. Clinical and Translational Impact Statement: The paper explores transfer learning’s impact on translating deep learning for medical image analysis, potentially boosting clinical impact. By leveraging pre-trained models and overcoming data limitations, it aims to improve accuracy, efficiency, andgeneralizability, paving the way for real-world medical applications.