A fully optimized deep learning framework for explainable and efficient kidney stone detection in computed tomography...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Kidney stone detection on axial computed tomography (CT) is critical for patient diagnosis and treatment.
• Existing deep learning models often require substantial computational resources.
• There is a need for efficient and accurate AI models for kidney stone detection.

Purpose of the Study:

• To develop and evaluate a hyperparameter-optimized custom convolutional neural network (CNN) for kidney stone detection on axial CT.
• To compare the custom CNN with widely used pretrained deep learning models.
• To emphasize statistical reliability, computational efficiency, and cross-institutional generalization.

Main Methods:

• A custom CNN with residual blocks was trained on 3,364 axial CT slices.
• Bayesian hyperparameter optimization was applied to the custom model and six pretrained backbones.
• Performance was evaluated using five-fold cross-validation and an independent external cohort; interpretability was assessed with Grad-CAM++.

Main Results:

• The custom CNN and ResNet101 achieved the highest test performance (accuracy 0.9960, F1 0.9957).
• The custom CNN demonstrated significantly lower inference latency and fewer parameters compared to baselines.
• Cross-validation showed robust performance, and external validation confirmed generalization capabilities.

Conclusions:

• Custom architecture search can yield models matching pretrained backbones' performance at a reduced computational cost.
• Consistent methodological benchmarking is essential for developing clinically deployable AI for kidney stone detection.
• The developed custom CNN offers an efficient and accurate solution for kidney stone detection.