Learning to Discover Explainable Clinical Features With Minimum Supervision.

Area of Science:

• Medical Imaging
• Machine Learning
• Computer Vision

Background:

• Limited labeled data poses a challenge for developing accurate deep learning models in medical imaging.
• Optical coherence tomography (OCT) is crucial for diagnosing various eye conditions.

Purpose of the Study:

• To compare the effectiveness of supervised transfer learning and semisupervised learning for in-depth knowledge acquisition with limited data in the OCT domain.
• To evaluate model performance using explainability techniques like local interpretable model-agnostic explanations and class activation maps.

Main Methods:

• Utilized EfficientNet-B4 for transfer learning and SimCLR for semisupervised learning.
• Trained models on a large public OCT dataset (108,312 images) and smaller subsets (31,200 and 4000 images).
• Employed local interpretable model-agnostic explanations and class activation maps for model interpretability.

Main Results:

• Supervised transfer learning with EfficientNet-B4 achieved 0.976 accuracy, 0.973 sensitivity, and 0.991 specificity on the limited dataset.
• Semisupervised learning with SimCLR (10% labeled data) achieved 0.946 accuracy, 0.941 sensitivity, and 0.983 specificity.
• Semisupervised learning demonstrated comparable performance to supervised methods using substantially less labeled data.

Conclusions:

• Semisupervised learning shows significant potential for OCT analysis, effectively utilizing unlabeled data with minimal labeled samples.
• This approach reduces the need for extensive data labeling, expertise, and time, making model development more efficient.