Non-small cell lung cancer (NSCLC), the most common type of lung cancer, is one of the deadliest malignancies globally. As precision management of lung cancer gains increasing attention, accurately identifying its histological subtypes is crucial for improving diagnostic and therapeutic outcomes. Traditional clinical diagnostic methods for NSCLC pathological subtypes are limited by their invasiveness, reliance on physician experience, and high consumption of medical resources. To address these challenges, our research team explores radiomics and deep learning technologies based on CT imaging to achieve non-invasive integrated diagnosis of lung cancer lesions, including detection, benign/malignant classification, and subtype classification.

In the detection of lung cancer lesions, we propose an enhanced Faster-RCNN method using Focal loss for early pulmonary nodule detection. This method adopts a two-stage architecture of candidate region detection and false-positive suppression, improving Faster-RCNN technology through 3D convolutional neural network (3D-CNN) techniques to detect early-stage lung cancer lesions. Experimental results demonstrate advanced performance on datasets such as LUNA16.

In benign/malignant classification, our team develops an intelligent pulmonary nodule diagnostic system that integrates 3D-CNN with multiple kernel learning support vector machines (SVM-MKL). By jointly analyzing CT imaging features and patient clinical information (e.g., age, smoking history, cancer history), the system enables multidimensional auxiliary diagnosis. Using a 34-layer 3D residual network (3D-ResNet), the system deeply extracts spatiotemporal features from CT images and dynamically fuses imaging features with clinical data through multiple kernel learning algorithms, addressing the challenge of co-optimizing heterogeneous feature spaces. Experiments show that joint modeling of imaging deep features and clinical information significantly enhances the diagnostic performance for benign/malignant differentiation of pulmonary nodules, validating the clinical value of multi-source data fusion strategies. Through an end-to-end automated analysis workflow, the system provides an efficient and interpretable intelligent solution for early lung cancer screening.

In the diagnosis of lung cancer subtypes, our team proposes a fusion model that combines radiomic features with deep learning features for comprehensive decision-making. By fully extracting deep features from CT images using 3D-CNN and quantifying features through radiomics, the model ultimately fuses the two types of features using a multi-head attention (MHA) mechanism for classification. This model represents the first study to integrate multi-source features and multimodal learning methods in the histological subtype classification of lung cancer. Experiments on an NSCLC dataset combining radiomics and radiogenomics demonstrate that the model achieves an accuracy of 0.88 and an area under the receiver operating characteristic curve (AUC) of 0.89 in distinguishing pulmonary adenocarcinoma (ADC) from squamous cell carcinoma (SqCC). These results validate the feasibility of predicting lung cancer histological subtypes using non-invasive imaging technologies.