Transfer Learning Techniques for Biomedical Image Analysis and Automated Disease Detection Systems

Abstract

Medical imaging is a valuable tool for disease detection, and advancements in diagnostic imaging technologies have enabled the early identification of various medical conditions. However, conventional manual image analysis approaches are time-consuming and often affected by intra- and inter-observer variability. These constraints can be circumvented through the use of automated medical image processing algorithms, which offer diagnostic solutions that are accurate, efficient, and consistent. This study presents a hybrid GRU+BiLSTM-based framework for biomedical image classification and automated disease detection using the NIH ChestX-ray14 dataset. The proposed approach incorporates image preprocessing techniques, including denoising, enhancement, augmentation, normalization, and data balancing, to improve image quality and model robustness. The suggested Hybrid GRU+BiLSTM model attained a 97.3% accuracy (ACC), 96.9% precision (PRE), 97.5% recall (REC), and 97.8% F1-score (F1), as shown in the experimental data. Results from comparisons with other methods demonstrate that the suggested model is superior than VGG16, ResNet50, DenseNet-121, and EfficientNet-B4. The results show that the suggested framework is useful for automated and dependable biomedical image-based disease identification.