Fine-tuned deep transfer learning: an effective strategy for the accurate chronic kidney disease classification.

Kidney diseases are becoming an alarming concern around the globe. Premature diagnosis of kidney disease can save precious human lives by taking preventive measures. Deep learning demonstrates a substantial performance in various medical disciplines. Numerous deep learning approaches are suggested in the literature for accurate chronic kidney disease classification by compromising on architectural complexity, classification speed, and resource constraints. In this study, deep transfer learning is exploited by incorporating unexplored yet effective variants of ConvNeXt and EfficientNetV2 for accurate and efficient classification of chronic kidney diseases. The benchmark computed tomography (CT)-based kidney database containing 12,446 CT scans of kidney tumor, stone cysts, and normal patients is utilized to train the designed fine-tuned networks. However, due to the highly imbalanced distribution of images among classes, the operation of data trimming is exploited for balancing the number of CT scans in each class, which is essential for designing an unbiased predictive network. By utilizing fine-tuned pre-trained models for our specific task, the training time is reduced leading to a computationally inexpensive solution. After the comprehensive hyperparameters tuning with respect to changes in learning rates, batch sizes, and optimizers, it is depicted that the designed fine-tuned EfficientNetV2B0 network of 23.8 MB in size with only 6.2 million architectural parameters shows substantial diagnostic performance by achieving a generalized test accuracy of 99.75% on balanced CT kidney database. Furthermore, the designed fine-tuned EfficientNetV2B0 attains high precision, recall, and F1-score of 99.75%, 99.63%, and 99.75%, respectively. Moreover, the final fine-tuned EfficientNetV2B0 ensures its scalability by achieving an impressive diagnostic accuracy of 99.73% on the test set of the original CT kidney dataset as well. Through the extensive evaluation of the proposed transfer learning strategy, it is concluded that the proposed design of fine-tuned EfficientNetV2B0 outperforms its counterparts in terms of accuracy and computational efficiency for chronic kidney disease diagnosis tasks. The final fine-tuned EfficientNetV2B0 serves as an accurate, efficient, and computationally inexpensive solution tailored for real-time deployment on medical or mobile edge devices.