Table 1 Summary of semi-supervised learning methods in medical image analysis
Method | Ref | Image modality | Metrics | Issues |
|---|---|---|---|---|
Self-Training semi-supervised learning | [4], 2023 | CT scan | Dice Similarity Coefficient (DSC), Intersection over Union(IoU), Precision, Recall | The model ’s effectiveness heavily relies on the initial teacher model ’s quality |
[5], 2024 | skin lesion images were obtained using dermoscopy | mIoU | The generalizability of the model to other datasets is lacking | |
[24], 2023 | X-ray | Accuracy, Precision, Recall, F1 Score | The method assumes that the distribution of labelled and unlabeled data can be aligned using small-paced self-training | |
Co-Training semi-supervised learning | [7], 2024 | MRI | DSC, HD | Potential computational complexity due to the co-training process, reliance on the initial quality of the models, and scalability issues for larger datasets |
[6], 2024 | CT scans | Accuracy, Precision, Recall, F1 Score | The quality and diversity of training data significantly impact the model’s performance. Sparse or noisy data can affect prediction accuracy | |
Graph-Based Semi-supervised learning | [11], 2023 | RGB | Accuracy, Precision, Recall, F1 Score, Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI) | The spectral decomposition process can be computationally intensive, which might limit the method ’s scalability for enormous datasets |
[10], 2024 | Synthetic Aperture Radar (SAR) images | accuracy, precision, recall, F1 score, and Specificity | The method ’s accuracy is sensitive to how well the similarity graph represents the underlying data structure | |
[9], 2009 | CT and MRI | accuracy, precision, recall, F1 score, and Specificity | Multi-label datasets often suffer from imbalanced label distributions, which can affect the performance of the graph-based semi-supervised learning methods | |
Pseudo-Labeling semi-supervised learning | [12], 2024 | 3D gadolinium-enhanced MR imaging scans (GE-MRIs | Dice Similarity Coefficient (Dice) | The effectiveness of the method relies on the accuracy of the initial pseudo-labels |
[13], 2023 | Pancreas-CT | Dice Similarity Coefficient (Dice) | The process of generating and refining soft pseudo-labels can be computationally expensive | |
Ensemble semi-supervised learning | [16], 2023 | colonoscopy and laryngoscopy CT | accuracy, precision, recall, F1 score, and Specificity | Ensembles with many models can still overfit the training data if not adequately regularized or if the base models are too complex |
[25], 2021 | Colon-CT | accuracy, precision, recall, F1 score, and Specificity | The model ’s effectiveness relies on the availability of high-quality annotated datasets, which can be time-consuming and costly to produce | |
Generative semi-supervised learning | [26], 2019 | RGB | accuracy, precision, recall, F1 score, and Specificity | Ensuring synchronization between the central server and distributed nodes can be challenging, especially as the number of nodes increases |
[27], 2019 | satellite images | Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR) | The effectiveness of the model on different types of cloud formations and satellite data | |
[28], 2024 | CT scans | Dice Similarity Coefficient (DSC), Intersection over Union (IoU) | The effectiveness of transfer learning depends on the quality and relevance of the pre-trained models used |