Improving adversarial learning for lung cancer tumor segmentation: The role of quasi-deep supervision framework with dual discriminators

Accurate segmentation of lung tumors in CT scans supports radiological assessment, diagnosis, and surgical planning. Although supervised learning methods achieve excellent performance, they rely heavily on large-scale annotated datasets. Methods based on Generative Adversarial Networks (GANs) alleviate annotation dependence but still lag behind supervised models. To overcome these challenges, this work proposed the Quasi-Deep Supervised Segmentation Generative Adversarial Network (QDS-SegGAN) model, which integrated deep-layer constraints into the GAN architecture. The generator, Edge Enhancement Network (EENet), utilized a combination of the Receptive Field Block (RFB) module and Parallel Encoder (PE) module to capture both global and local contextual information. An Edge Enhancement Attention (EEA) module preserved critical boundary details, while the Edge Loss (EL) function refined boundary accuracy. The model incorporated two discriminators, an anatomy-aware discriminator (D1) and a boundary-refinement discriminator (D2), to provide multi-level guidance for the generator’s decoder outputs. The proposed method was evaluated on both the private FPHC-LC dataset and the public LIDC-IDRI dataset, where it consistently outperformed existing supervised and GAN-based segmentation methods. Specifically, QDS-SegGAN achieved an Intersection over Union (IoU) of 69.80%, a Dice coefficient of 81.59%, and a Hausdorff Distance (HD) of 2.55 mm on the FPHC-LC dataset, demonstrating its effectiveness and strong generalization capability. Overall, the proposed framework improves segmentation accuracy under limited annotation conditions, providing reliable tumor delineation for computer-assisted lung cancer assessment.