SISE-PC: Semi-supervised Image Subsampling for Explainable Pathology Classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Deep learning (DL) models for automated pathology classification are effective but require substantial data and computational resources.
• Optical coherence tomography (OCT) imaging generates complex data that benefits from automated analysis.

Purpose of the Study:

• To decrease the data and computational costs associated with training DL models for OCT image classification.
• To introduce an active learning framework for efficient OCT image dataset selection and model fine-tuning.

Main Methods:

• Pre-training a ResNet feature extractor using SimCLR contrastive loss for latent encoding of OCT images.
• Developing an active learning framework utilizing label propagation on latent encodings to identify uncertain OCT image samples.
• Fine-tuning the pre-trained ResNet model on a minimal, sub-sampled dataset and visually explaining pathological sites.

Main Results:

• The proposed framework identified a minimal subset (as low as 2%) of uncertain OCT images requiring specialist attention.
• Fine-tuning on this minimal dataset achieved up to 97% classification accuracy for OCT image analysis.
• The method demonstrated significant reductions in data and compute requirements for DL training.

Conclusions:

• The active learning framework effectively reduces the cost of DL model training for OCT image classification.
• This approach enables efficient identification of critical samples, enhancing diagnostic capabilities.
• The methodology shows promise for extension to other medical imaging modalities to minimize prediction costs.