Enhancing Two-Year Recurrence-Free Survival Prediction in Non-Small Cell Lung Cancer (NSCLC) Patients Using Tumor-Centric Attention Network (TCA-Net)

Accurate predictions of postoperative recurrence are essential for determining appropriate follow-up treatments after surgery, as patients with non-small-cell lung cancer (NSCLC) at the same clinical stage have different recurrence incidences. However, simple convolutional neural network (CNN)-based methods are limited when presented with tumors of various sizes. This study aims to predict two-year recurrence-free survival precisely in patients with tumors of various sizes on preoperative CT images using what is termed a Tumor-Centric Attention Network (TCA-Net).

The proposed network features dual branches, each with an identical architecture but distinct weights to extract diverse features from CT images and tumor masks simultaneously. The tumor-centric attention module integrates two disparate feature maps at each level to amplify the characteristics of the tumor. All feature maps are concatenated with the finest resolution, enabling the extraction and integration of comprehensive multi-scale features for the complex tumor environment.

TCA-Net showed an accuracy of 75%, balanced accuracy of 75.05%, specificity of 76.16% and an AUC value of 0.78. These results represent more balanced accuracies by 4.76% and 2.58% compared to ResNet-18 with CT images and dual ResNet-18s with CT images and tumor masks, respectively. Specifically, TCA-Net demonstrated a substantial improvement in the small-sized tumor group, achieving a balanced accuracy of 81.32%, sensitivity of 85.71%, and specificity of 76.92%.

TCA-Net improved the prediction performance of two-year recurrence-free survival on average across tumors of all sizes, with significant improvements, especially for small-sized tumors.