Lung cancer is a critical health issue, and early detection is crucial for enhancing patient outcomes. This study presents a novel framework for generating three-dimensional (3D) representations of lung tumors from computed tomography (CT) scans, addressing three key challenges in the analysis process. Firstly, we address the precise segmentation of lung tissues, which is complicated by a high proportion of non-lung pixels that skew the classifier. Our method uses a customized generative adversarial network (GAN) enhanced with an off-policy proximal policy optimization (PPO) strategy. This strategy enhances segmentation performance by addressing inherent classifier biases and implementing a reward system to more accurately identify minority samples. Secondly, the framework enhances tumor detection in the segmented areas by employing a specialized GAN trained with an adversarial loss, which helps the generator create tumor regions that match real ones in both shape and internal features, even when contrast is low or boundaries are unclear. Thirdly, after tumor detection, the EfficientNet model extracts essential features for 3D reconstruction. The features are then enhanced by a spatial attention-based transductive long short-term memory (TLSTM) network for better performance. The TLSTM network enhances performance by assigning greater weight to samples near the test point within a transductive learning framework. Tested on the Lung Image Database Consortium Image Collection (LIDC-IDRI) dataset, our methodology achieved Hausdorff distance (HD) and Euclidean distance (ED) metrics of 0.648 and 0.985, respectively, indicating superior performance compared to existing methods. Our research introduces a clinical tool that significantly boosts the capabilities of radiologists in diagnosing and planning treatment for lung cancer. Code is publicly available at https://github.com/ZhisenHe/3D-representation/ .