Table 2 Comparison of advanced models for lung segmentation.
Authors | Methodology | Contribution | Dataset | Result | Limitation |
|---|---|---|---|---|---|
Gite et al.28 | U-Net++ for lung segmentation | Enhances TB detection with advanced segmentation | 138 X-ray images | IoU: 0.95 | Specific to X-ray images, not other modalities |
Arvind et al.29 | Modified U-Net with dropouts | Reduces overfitting, enhances model generalization | 900 X-ray images | Accuracy: 93.87% | Optimized for lung segmentation, less versatile for other organs |
Chen et al.30 | LDANet combines RSA and GCA | Enhances lung CT segmentation via dual attention | LIDC-IDRI | Dice similarity coefficient (DSC): 98.430% | Sensitive to variations in imaging protocols |
Ghali et al.31 | Dual loss functions with five models | Optimizes CXR segmentation for various lung conditions | 662 X-ray images | F1 score: 97.47% | May not generalize across all imaging equipment types |
Murugappan et al.32 | DeepLabV3 + with various networks | Enhances segmentation of lung and infected areas | 750 CT images | IoU: 0.9971 | Performance varies with different network layers |
Ambesange et al.33 | FL with U-net and transfer learning | Enhances lung X-ray segmentation while preserving data privacy | 662 X-ray images | Accuracy: 98.92% | Limited by data variability across nodes |
Hasan et al.34 | Deeplabv3plus with Atrous Convolution | Optimizes feature resolution for segmentation | 558 X-ray images | Accuracy: 97.42% | May struggle with highly variable image quality |
Swaminathan et al.35 | Wiener filter, GAN with SSSOA, VGG16 classification | Streamlines detection process with an integrated model | LIDC-IDRI | Accuracy: 97% | Relies on high-quality pre-processed images |
Suji et al.36 | U-Net with EfficientNet-b3, transfer learning | Optimizes nodule segmentation with transfer learning | LIDC-IDRI | IoU: 0.45 | Performance variability across datasets |
Thangavel and Palanichamy37 | T-Net, CenterNet, NASNet with preprocessing | Automates nodule classification efficiently | LIDC-IDRI | DSC: 99.07% | May require fine-tuning for new datasets |
Cai et al.38 | GANs for image translation segmentation | Enhances lung CT segmentation with GANs | 267 CT images | Accuracy: 89.63% | May not handle all types of lung anomalies |
Ramos and Pineda38 | Tri-phase semi-automated segmentation with preprocessing | Provides consistent, swift segmentation results | 267 CT images | IoU: 0.9341 | May require manual adjustments for optimal results |
Zheng et al.40 | Threshold-gradient with TFDM and convex hull repair | Enhances segmentation performance and robustness | 2112 CT images | IOU: 0.9911 | Requires precise calibration of threshold settings |
Guo et al.41 | whale optimization algorithm-based random mutation strategy for image segmentation | Speeds up convergence, improves segmentation | 25,000 CT images | Feature similarity index measure (FSIM): 81.57% | May not adapt well to other cancer types |
Pandey and Bhandari42 | Morphological filtering and SVM classification | Enhances tumor visibility for early intervention | LIDC-IDRI | Accuracy: 87.79% | May miss micro-tumors or highly diffuse anomalies |
Vijayakumar et al.43 | CapsNet with RGF and SMFMF, U-net segmentation | Enhances lung cancer detection efficiency | 1097 CT images | Accuracy: 98% | May underperform with non-standardized data |
Qiao et al.44 | WSTSA with genetic algorithm | Streamlines optimal CNN architecture selection | 2773 CT images | DSC: 72.40% | Dependent on initial algorithm parameters |
Li et al.45 | Attention and Sobel edge detection | Enhances lesion feature extraction and performance | LIDC-IDRI | Accuracy: 93.48% | May require high computational resources |
Murugaraj et al.46 | HM-LeNet with NLM filter and K-Net | Classifies lung abnormalities efficiently | 23 CT scans images | Accuracy: 92.7% | Potential overfitting to specific abnormalities |
Fu et al.47 | LDN-SNP with SNP-type neurons | Efficient, compact segmentation of COVID-19 CT | 100 CT scans images | Accuracy: 97.0% and DSC: 74.0% | Limited adaptability to other imaging tasks |
Erciyes et al.48 | ANN with optimized U-Net and augmentation | Reduced overfitting with enhanced segmentation | American association of physicists in medicine (AAPM) thoracic auto-segmentation | DSC: 94.68% and IoU: 0.8990 | Limited adaptability to unseen pathologies |
Yang et al.49 | 3D model with focal and dice loss | Improved deep and shallow feature representation | 199 CT scans images | Accuracy: 99.90% and DSC: 56.10% | Overfitting risk from a small COVID dataset |
Nguyen et al.50 | Two-stage no-new-U-Net with attention and boundary loss | Improved GGO segmentation with fuzzy boundary handling | Post-COVID CT challenge | DSC: 71.93% | Difficulty generalizing to non-GGO lung lesions |
Vincy et al.51 | 3D ResNet50 encoder with dense-feature U-Net decoder | Enhanced multiscale tumor features from lung CT slices | LUNA16 | IoU: 0.738 | Contrast issues in small or diffuse tumors |
Jannat et al.52 | Lightweight residual U-Net with CBAM and ASPP | Improved lung mask extraction | Japanese society of radiological technology (JSRT) | DSC: 98.72% | Limited generalization on low-quality images |
