Enhanced CT image classification for kidney stones using pruned ConvNeXt and two-tier optimization

In this work, an innovative two-tier feature-optimized deep learning model is proposed for kidney stone detection in CT images. The framework uses a new modified ConvNeXt learning model to capture complex image patterns from CT images, along with the use of a LightGBM classifier.

A two-tier optimization strategy is implemented to refine the detection process. Initially, Dynamic Channel Pruning is used within the ConvNeXt architecture to identify and retain the most informative channels during feature extraction. By dynamically evaluating the importance of each channel, this step processes the most relevant channels, reducing computational complexity and highlighting critical features. Then, the Pufferfish Optimization Algorithm (POA) is applied for optimal feature selection. This optimization helps isolate the most discriminative features for kidney stone detection. Additionally, POA is applied to adjust the hyperparameters of the LightGBM classifier to increase classification accuracy and efficiency.

The proposed model achieves the highest accuracy of 97.8%, compared with other models.

The proposed model achieves enhanced detection accuracy and efficiency through the use of Dynamic Channel Pruning and Pufferfish Optimization. This model offers a promising solution for kidney stone detection in CT images.