A CNN-transformer fusion network for predicting high-grade patterns in stage IA invasive lung adenocarcinoma.

Abstract

Background: Invasive lung adenocarcinoma (LUAD) with the high-grade patterns (HGPs) has the potential for rapid metastasis and frequent recurrence. Therefore, accurately predicting the presence of high-grade components is crucial for doctors to develop personalized treatment plans and improve patient prognosis.

Purpose: To develop a CNN-transformer fusion network based on radiomics and clinical information for predicting the HGPs of LUAD.

Methods: A total of 288 lesions in 288 patients with pathologically confirmed invasive LUAD were enrolled. Firstly, radiomics features were extracted from the entire tumor region on lung computed tomography (CT) images and then fused with clinical patient characteristics. Secondly, a structure was proposed that concatenated a convolutional neural network (CNN) and Transformer encoding blocks to mine and extract more comprehensive information. Finally, a classification prediction was performed through fully connected layers.

Results: Accuracy, sensitivity, specificity, precision, and area under the receiver operating characteristic (ROC) curve (AUC) were utilized for evaluation of the model's classification prediction performance. Delong's test was used to compare the AUCs of different models for significance. The proposed model was effective with an accuracy of 0.86, sensitivity of 0.67, specificity of 0.94, precision of 0.74, and AUC of 0.91.

Conclusions: The CNN-transformer fusion network, based on radiomics and clinical information, demonstrates good performance in predicting the presence of HGPs and can be employed to assist in the development of personalized treatment plans for patients with invasive LUAD.