EfficientNetB3-HCAF: A Hybrid Deep Learning Framework for Chest Cancer Classification Using Chest CT Images

Abstract

Chest cancer is among the common causes of death attributed to cancer globally, where early diagnosis of the disease plays an important role in increasing survival rates. Deep learning-based methods have made remarkable improvements in the automatic analysis of medical images. Nevertheless, some of the existing solutions have some drawbacks associated with the representations of multiscale lesions, dependency learning, and high computational costs. In this regard, the present study has introduced a new hybrid deep learning method called EfficientNetB3-HCAF to automatically classify patients with chest cancer based on the images taken from chest CT scans. Specifically, the architecture used for the model is based on pre-trained EfficientNetB3 backbone and a novel Hybrid Contextual Attention Fusion (HCAF) module. The HCAF module is comprised of several blocks such as multi-scale depthwise-separable convolution, Convolutional Block Attention Module (CBAM), transformer encoder, and residual refinement. In addition to training, extensive preprocessing and data augmentation were performed to increase the generality of the model. From the experiment results, it can be concluded that the model proposed in this study shows reliable convergence and classification performance by achieving accuracy of 95%-96%. Results from the confusion matrix analysis further prove that there have been fewer instances of false positives and negatives, proving that the diagnostic accuracy is indeed reliable. Moreover, the efficient architecture minimizes computational difficulty while ensuring robust feature extraction capabilities. Thus, the EfficientNetB3-HCAF system presents an effective and computationally efficient solution for computer-aided diagnosis of chest cancer.