Cholecystitis detection based on PoolConvNeXt using computed tomography images

CNNs, though less preferred than transformers in computer vision, hold significant potential in image classification and segmentation. Inspired by PoolFormer, we propose PoolConvNeXt, a novel CNN model. To evaluate its performance, we used a cholecystitis dataset and a large cancer dataset with 60,000 images across 12 classes. PoolConvNeXt features four stages: stem, main, downsampling, and output. The main stage integrates pooling and Kolmogorov Arnold Network-like structures for feature extraction. Using the pretrained PoolConvNeXt, we design a Deep Feature Engineering (DFE) model with three phases: feature extraction via overlapping patches, feature selection using Cumulative Weighted Neighborhood Component Analysis (CWNCA), and classification with shallow classifiers. PoolConvNeXt achieved 97.99% validation accuracy and 97.23% test accuracy on the cancer dataset. On the cholecystitis dataset, the pretrained DFE model attained over 90% accuracy, with kNN and SVM reaching 100%. These results demonstrate the proposed model’s strong capability in biomedical image classification.