A two-stage deep learning framework for kidney disease detection using modified specular-free imaging and EfficientNetB2

Kidney diseases represent a substantial public health concern, with their incidence increasing markedly over the past decade. Addressing this challenge, our research introduces a sophisticated two-stage diagnostic model for enhancing the detection accuracy of various kidney pathologies. The initial stage of the proposed model comprises a novel Modified Specular-Free (MSF) technique designed to improve the visual quality of renal images. This technique adaptively enhances image details by applying targeted enhancements for more discrimination between dark and bright luminance levels. The objective of this technique is to restore critical image information and enrich color representation in diagnostically-significant areas. The enhanced images are then processed through the second stage, which involves classification using the EfficientNet-B2 deep learning architecture. Our model was rigorously compared against a suite of established pre-trained models, including VGG16, ResNet50, VGG19, DenseNet121, DenseNet169, DenseNet201, EfficientNet-B0, EfficientNet-B1, and EfficientNet-B3. Comprehensive testing revealed that our model not only outperforms these benchmarks, but does so with a notable accuracy of 98.27%. The robustness of the model was further ensured through its capability to effectively differentiate between normal renal conditions and various pathologies such as tumors, kidney stones, and cysts. This research not only demonstrates the potential of integrating advanced image enhancement techniques with cutting-edge classification models but also introduces a scalable approach for improving diagnostic accuracies in other complex medical imaging contexts.