Table 13 Comparative analysis of the proposed lung segmentation, tumor detection, and 3D reconstruction models against established models.

Lung segmentation

Our introduced approach achieved an accuracy of 92.46%, IoU of 0.881, and an HD of 1.152, markedly surpassing current models, including U-Net-ANN-AUG⁴⁸, 3D-Lightweight-AttnROI⁴⁹, 2D3D-Attn-Boundary⁵⁰, and 3D-ResNet-UDecoder⁵¹. This advancement is evidenced by substantial improvements in segmentation accuracy and reductions in HD, showcasing enhanced quality in contour detection

Accuracy:12.26%, IoU: 8.5%, HD: 33.5%

Employing off-policy PPO strengthens the ability of the model to handle classifier imbalance efficiently, enhancing clarity in delineating lung tissues

Tumor detection

Relative to competing frameworks such as MSDA-BiFPN⁷³ and GCSAM-CNDNet⁷⁴, our approach yields higher accuracy and IoU and reduced HD metrics (accuracy = 93.45%, IoU = 0.893, HD = 0.721), underscoring its improved proficiency in delineating tumor margins

Accuracy: 9.18%, IoU: 7.85%, HD: 5.63%

Improved quality in detecting tumors stems from implementing a new loss function in the GAN framework, which refines the demarcation of tumor edges, outperforming conventional deep learning techniques

3D reconstruction

The model outperforms others in the 3D reconstruction, achieving the highest accuracy and smallest HD and ED (accuracy = 90.84%, HD = 0.648, ED = 0.985) compared to techniques such as ViT⁸⁸ and several GAN-based methods. This highlights the performance of our proposed model in generating highly detailed and precise 3D representations of tumor structures

Accuracy:17.26%, ED: 22.71%, HD: 43.92%

The quality of 3D reconstructions has greatly improved through the use of advanced methods for lung segmentation and tumor identification, augmented by the refined temporal analysis features of TLSTM networks