Reinforcement learning-driven imbalanced classification enhanced by a GAN model augmentation for lung cancer detection.
Authors
Affiliations (3)
Affiliations (3)
- Department of General Surgery, The First Affiliated Hospital of Zhejiang Chinese Medical University (Zhejiang Provincial Hospital of Chinese Medicine), Hangzhou, Zhejiang, China.
- Emergency and Critical Care Center, Intensive Care Unit, Zhejiang Provincial People's Hospital (Affiliated People's Hospital), Hangzhou Medical College, Hangzhou, Zhejiang, China.
- Department of Intravenous Therapy Nursing，Nursing Department, Zhejiang Provincial People's Hospital (Affiliated People's Hospital), Hangzhou Medical College, Hangzhou, Zhejiang, China. [email protected].
Abstract
Lung cancer has been one of the largest causes of cancer related deaths globally with approximately five million deaths annually related to the condition. The capability of computed tomography (CT) scans to identify early and correctly can result in improved patient survival, yet the existing deep learning solutions suffer due to three remaining issues: the imbalance of the training sample, sensitivity to the choice of hyperparameters, and the lack of generalizability across clinical environments problems that have not been sufficiently addressed in the prior research. To overcome classification shortcomings, the study suggests a deep learning model that combines three essential elements: off-policy proximal policy optimization algorithm to tackle the class imbalance issue by over-rewarding correct minority classifications; online data augmentation approach based on a generative adversarial network with a new regularization to increase training diversity; and a better differential evolution algorithm involving k-means clustering and human mental search strategies to automatically optimize hyper. It is tested on two benchmark datasets LIDC-IDRI and LUNA-16 with five-fold stratified cross-validation. Our model attains F-measures of 90.128 and 92.547, respectively, surpassing eight state of the art baseline methods. These improvements are confirmed by statistical testing of significance. These findings show that the proposed framework has the potential to serve as a promising clinical decision-support tool for the early detection of lung cancer, pending further prospective multi-center validation.