OncoSeg2D: A deep framework for semantic segmentation of lung cancer in 2D CT scans.
Authors
Affiliations (4)
Affiliations (4)
- Senior Department of Oncology, Chinese PLA General Hospital, Beijing, China.
- Department of Oncology, Beijing Chaoyang Integrative Medicine Rescue and First Aid Hospital, Beijing, China.
- Department of Ultrasound Diagnosis, The Sixth Medical Center of Chinese PLA General Hospital, Beijing, China.
- Department of Integrated Traditional Chinese and Western Medicine, Affiliated Hospital of Hebei University, Baoding, Hebei, China.
Abstract
Lung cancer lesion segmentation in two-dimensional computed tomography (2D CT) images remains challenging due to blurred boundaries, heterogeneous morphologies, and annotation uncertainty, leading to unreliable delineations and reduced clinical usability. To address this research gap, we propose a novel 2D CT lung cancer semantic segmentation framework, OncoSeg2D, which explicitly tackles boundary ambiguity and morphological distortion through two complementary modules. Specifically, an Uncertainty-aware Boundary Modeling (UBM) module probabilistically represents tumor edges via learnable mean-variance estimation and gradient-weighted sampling, while a Morphology-Preserving Regularization (MPR) module constrains the segmentation with curvature, compactness, and convexity priors to maintain global shape consistency. The framework integrates these designs with multi-scale feature extraction from a SegFormer backbone and requires no additional annotations or three-dimensional (3D) reconstruction. Experiments conducted on the Medical Segmentation Decathlon Challenge dataset and the lung cancer segmentation dataset demonstrate that OncoSeg2D achieves IoU scores of 0.865 and 0.788, mIoU scores of 0.881 and 0.799, and Dice Similarity Coefficients (DSC) of 0.923 and 0.816, consistently outperforming conventional CNN-based models and mainstream Transformer-based methods. Compared with the SegFormer baseline, the proposed method improves mIoU by 3.8% and 4.0% on the two datasets, respectively, while reducing the Hausdorff distance from 5.61 to 3.41 and from 7.26 to 5.28, indicating superior boundary refinement and stronger global shape consistency. These results verify that explicitly integrating uncertainty modeling and morphological priors yields both higher accuracy and better interpretability. Overall, the proposed framework not only enhances segmentation accuracy but also improves clinical interpretability and reliability, offering a promising solution for lung cancer diagnosis assistance and therapeutic outcome monitoring.