Uncertainty-Aware Volumetric Transformer with Dual Spatial-Channel Attention for Lung Nodule Classification

Lung cancer is also among the most common causes of cancer-related deaths in the world, and the earliest possible detection of the cancer through computed tomography (CT) is important in the enhancement of patient survival. Nevertheless, accurate diagnosis is still a challenge as the nodules are small and indistinct, inter-rater consistency among radiologists, and the traditional deep learning systems have limited capacity to handle volumetric interactions and give interpretable and confidence-aware forecasts. This research suggests an uncertainty-cognizant Transformer-Enhanced Dual-Level Attention Network (TDA-Net) to classify lung nodules in CT images to deal with these issues. The suggested architecture combines a 3D Swin Transformer backbone and sequential spatial and channel attention fusion to be able to model both localized structural and global volumetric context. Moreover, Monte Carlo dropout is used in inference to measure predictive uncertainty and allows low-confidence cases to be identified and sent to a radiologist. The model is tested on a publicly available lung CT dataset, and it has an accuracy of 98.3% with high sensitivity to small nodules in the feature space. There is a separation of classes in the feature space, and the uncertainty rate is 5.1%. The findings of the experiment indicate that TDA-Net can be used as a supportive decision-making tool to diagnose lung cancer with the help of computers because it has better discriminative performance and uncertainty awareness when compared to the baseline models. Moreover, distinguishable uncertainty of predictions and uncertainty of models are present. Predictive uncertainty is measured through the variance of softmax probability distributions through stochastic forward passes, which is related to the ambiguity of data. Monte Carlo dropout is used to estimate model uncertainty as a Bayesian approximation, which represents parameter-level uncertainty due to a small amount of training data.