Evaluating CoAtNet for multiclass lung cancer classification on CT Images: a benchmark study on the IQ-OTH/NCCD dataset

Abstract

Accurate classification of lung cancer from Computed Tomography (CT) scans remains a significant challenge in medical image analysis. In this research, we introduce an incremental deep learning framework that leverages CoAtNet's hybrid convolution-attention architecture for optimal multiclass lung cancer classification. Our approach integrates extensive preprocessing, including class-balancing via selective augmentation, high-resolution resizing, and adaptive color-space transformations to mitigate dataset imbalance and improve generalizability. We fine-tuned a pretrained CoAtNet-0-RW-224 model on a selected subset of the IQ-OTH / NCCD data set, obtaining an overall classification accuracy of 98.17% and weighted precision of 97.81 %, recall of 97.77 %, and F1 score of 97.12 %. A comparative study through a confusion matrix and classification report confirms competitive performance in all three diagnostic categories: benign, malignant, and normal. Experimental results establish CoAtNet as a robust architecture for CT-based lung cancer screening and provide a compelling baseline for future research utilizing transformer-convolutional hybrid models in medical imaging applications.