Feature Copy-Paste with Fine-Grained Data Augmentation for Robust Lung Cancer Analysis in CT Images

With chest computed tomography (CT) serving as the primary imaging modality for lung cancer analysis, many diagnostic tasks that influence clinical decision-making can essentially be formulated as CT-based image classification tasks. For instance, non-invasive EGFR gene mutation prediction is a critical task, as EGFR status directly guides targeted therapy and affects prognosis. However, the genotype-relevant information in CT images is inherently weak and noise-sensitive, especially under cross-scanner and cross-parameter heterogeneity. Consequently, a robust and generalizable classification framework remains limited. To address this challenge, this paper proposes a novel general framework (F $_{}^2$ Mix) for CT-based lung cancer analysis, which integrates fine-grained data augmentation and feature copy-paste operations to achieve robust and comprehensive feature learning and refinement. Specifically, a Fine-Grained Augmentation (FGA) module is introduced to enhance both the textural and semantic information of the images. In parallel, a Feature Copy-Paste Consistency (FCPC) module imposes constraints in the feature space to ensure consistency between the image-level and feature-level representations. Furthermore, to mitigate feature redundancy during the final feature fusion stage, a Feature Refinement (FR) module is designed to effectively filter out redundant information from the network. The proposed method is a plug-and-play general framework that demonstrates significant improvements across six baseline methods on three important lung cancer analysis tasks, including lung nodule malignancy prediction, distant metastasis prediction, and EGFR gene mutation prediction, achieving AUC improvements ranging from 0.51% to 12.7% across different baselines and tasks. Extensive ablation and comparative experiments further validate the superior accuracy and robustness of the proposed approach. The source codes are available at https://github.com/499huangxingyu/F2Mix .