Hybrid CNN-Transformer Architecture for Early Lung Cancer Diagnosis from CT scans

Source abstract

The timely diagnosis of lung cancer is important in enhancing survival rate among patients who have the disease but the use of traditional diagnostic techniques cannot always determine the presence of minute pulmonary nodules during CT scan. The paper proposed a hybrid machine learning system based on Convolutional Neural Networks (CNNs) and Transformers to diagnose early lung cancer automatically on the basis of the IQ-OTH/NCCD CT scan dataset. The CNN element isolates local fine-grained features on the images whereas the Transformer element isolates long-range global relationships and a blend feature union mechanism joins these representations to end up with final classification of normal, benign, or malignant. Normalisation, resizing, noise removal and augmentation of the dataset were done to enhance robustness of the model. The hybrid CNN-Transformer model was compared to the basic models such as CNN alone and Vision Transformer (ViT) models. Experimental results show that the hybrid model achieved an overall accuracy of 93.7%, with class-wise precision, recall, and F1-scores of 95.1%, 94.7%, and 94.9% for normal, 91.2%, 90.5%, and 90.8% for benign, and 94.9%, 95.3%, and 95.1% for malignant cases. The confusion matrix showed that few misclassifications were experienced with benign and adjacent classes, and ROC analysis provided AUC values of 0.978, 0.942, and 0.981 with normal, benign, and malignant, respectively, to prove that there was an excellent discriminative performance. Such findings suggest hybrid CNN-Transformer architecture is effective in combining local and global features to produce reliable early detection of lung cancer. The model is more effective compared to conventional CNN and Transformer methods and is highly interpretable, which is why it has high potential to assist radiologists in making clinical decisions. The paper presents the opportunity of hybrid deep learning designs in terms of enhancing diagnostic quality and patient outcomes, and the future research will involve the use of 3D CT information, multimodal imaging, or application in the clinical environment.