Multiclass lung cancer detection using a hybrid capsule inspired deep neural network

Convolutional Neural Networks are widely used in lung cancer detection for more than a decade. However, it suffers from preserving spatial relationships among features, leading to dead units in deeper layers. This is overcome by the capsule network (CapsNet), which estimates various instantaneous parameters. Nevertheless, the dynamic algorithm implemented in CapsNet is prone to computational complexity because of its higher-end matrix multiplication between primary and secondary capsules. In this study, a light weight attention (LWA)-based EfficientNetB0 and Capsule-inspired feature encoding module is proposed to reduce computational complexity. The role of the LWA lies in its filtering capabilities, thus strengthening the important features and reducing the redundancy. This helps the proposed architecture work more effectively without requiring the computationally intensive routing algorithm. The proposed model targets multiclass classification of benign, normal, and malignant computed tomography images. Unlike CapsNet, no decoder network is present in the proposed architecture. Moreover, a simplified matrix multiplication is computed, which results in fewer floating point operations (FLOPs) of 0.01 GFLOPS. Although the proposed model attained 100% F1 score, accuracy, precision, and recall on the test set, the experiments proved that there is no data leakage. These results indicate its potential to support radiologists in lung cancer detection, though its real-world utility requires prospective clinical evaluation.