Parallel spatial cross attention ensembled deep learning model for lung cancer diagnosis and classification using computed tomography images

Abstract

Lung cancer is considered one of the deadliest cancer types, which affects the respiratory system severely. Globally, the rate of morality is increasing every year due to the lack of diagnosis of lung cancer at its early stage. Thus, lung cancer detection and classification are essential to decrease mortality and increase the probability of patient survival. However, the traditional lung cancer classification techniques possessed the drawbacks of poor generalization, high false positives, higher computational complexities, higher memory consumption and so on. Hence, this research proposes a Parallel Spatial Cross Attention ensemble Bidirectional long short-Term Memory (PSCABTM) model to classify lung cancer by overcoming the shortcomings of the existing models. The Parallel Spatial Cross Attention (PSCA) mechanism can capture the fine-grained features and long-range dependencies, which leads to accurate classification results. Moreover, the extraction of Deep Radiomics Transformer Features (DRTF) minimizes the computational complexity, resulting in lower memory usage and a faster training process. Additionally, the PSCABTM model learns hierarchical features in both forward and backward directions, which facilitates overcoming the limitation of traditional techniques and leads to better classification. According to the experimental results, the proposed model outperforms existing methods and attains F1-score, accuracy, specificity, recall, and precision of 99.20%, 99.20%, 98.60%, 99.76% and 98.15% using the LUNA16 dataset, respectively.