Lung cancer is one of the most dominant causes of cancer-related mortality globally where the early and accurate lung nodule classification based on computed tomography (CT) scan is vital in improving the survival rate. However, the time-consuming nature of manual interpretation of CT images and inter-observer variability creates a need for reliable automated classification systems. Current deep learning methods are limited in their integration of multi-scale contextual features, modelling of channel-spatial dependencies, and interpretability, which limits their clinical applicability. In this project, we introduce LungCare-AI, a deep learning-based system for classifying lung nodules from CT images, designed to improve feature representation and interpretability. The key model, SE-AttendNet, combines three complementary components: (i) multi-scale patch-based feature extraction using dilated convolutions to capture the variability of lesions, (ii) squeeze-and-excitation (SE) blocks to recalibrate channels adaptively, and (iii) multi-head spatial attention to model non-local dependencies across regions. The proposed architecture, in contrast to traditional sequential attention designs, is jointly refined across both channel and spatial dimensions to enhance discriminative learning. Experiments on the publicly available LIDC-IDRI dataset show that the proposed model achieves 93.6% accuracy, 91.7% F1-score, and 94.6% AUC, outperforming a series of baseline architectures under consistent experimental conditions. Ablation analysis is used to verify the complementary role of multi-scale representation, channel attention and spatial attention. Moreover, Grad-CAM-based visualisations can provide qualitative insight into the model's decisions, highlighting salient regions associated with the predicted classes. Although the results imply better classification performance and improved interpretability, the framework is tested on a single dataset. It relies on ROI-based preprocessing, which may limit its generalisation across different clinical settings. Future directions will include cross-dataset validation and more quantitative measures of explainability. Overall, LungCare-AI provides an interpretable framework for ROI-based lung nodule classification that may support future computer-aided clinical decision-support applications.
Rajkumar et al. (Tue,) studied this question.