Learning With Fewer Images via Image Clustering: Application to Intravascular OCT Image Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning excels at medical image segmentation across various modalities.
• Manual annotation is time-consuming, challenging, and requires skilled analysts.
• Reducing annotated data is crucial for efficient deep learning model development.

Purpose of the Study:

• To develop and evaluate methods for reducing annotation effort in medical image segmentation.
• To compare deep learning-based clustering with equally-spaced subsampling for intravascular optical coherence tomography (IVOCT) image annotation.
• To assess the impact of sampling strategies on coronary calcification segmentation accuracy.

Main Methods:

• Tested two annotation reduction schemes: equally-spaced image subsampling and deep learning-based clustering.
• Clustering involved autoencoder feature extraction, k-medoids clustering, and using cluster medians for training.
• Compared U-net model performance trained on full datasets versus subsampled datasets for coronary calcification segmentation.

Main Results:

• The deep learning-based clustering approach outperformed or matched equally-spaced subsampling across tested ratios.
• With only 10% of images, clustering improved the mean F1 score for calcific class from 0.52 to 0.63.
• Sampling images from more volumes of interest (VOIs) improved performance for a fixed number of training images.

Conclusions:

• Recommends the clustering-based approach for annotating a small fraction of IVOCT images to create baseline models.
• Suggests that further improvements can be achieved by incorporating active learning strategies for image selection.
• Highlights the potential of reduced annotation strategies for efficient medical image segmentation.