Deep Transfer Learning with Customized Convolutional Neural Networks for Radiological Image Classification

Radiological image interpretation is fundamental to modern clinical diagnosis, yet remains challenging due to high data dimensionality, complex anatomical variability, and the limited availability of labeled datasets. This study proposes a deep learning framework that integrates a customized convolutional neural network (CNN) with ResNet50- based transfer learning to improve diagnostic accuracy and robustness in radiology. The bespoke CNN is designed to capture hierarchical spatial representations from medical images, while ResNet50 leverages pretrained weights to accelerate convergence, preserve generic visual features, and enhance generalization to domain-specific radiological patterns. Extensive experiments conducted on multiple radiology datasets demonstrate that the proposed approach consistently outperforms CNNs trained from scratch, achieving accuracy exceeding 91% alongside notable improvements in Precision, Recall, F1-score, and Area Under the Curve (AUC). Gradient-weighted Class Activation Mapping (Grad-CAM) further reveals that the model focuses on clinically relevant regions such as lesions and abnormal tissue structures, thereby providing interpretable predictions and strengthening clinical trust. The results confirm that transfer learning significantly mitigates overfitting, stabilizes model training, and delivers reliable performance in limited-data scenarios commonly encountered in medical imaging. In addition, the proposed framework exhibits faster convergence and improved consistency across heterogeneous datasets, highlighting its robustness and adaptability. Overall, this work presents a scalable and explainable solution for automated radiological diagnosis, demonstrating how the integration of pretrained deep architectures with task-specific CNN design can advance AI-assisted clinical decision support and contribute toward more accurate, efficient, and reliable medical image analysis.