ENRICH: Exploiting Image Similarity to Maximize Efficient Machine Learning in Medical Imaging

Abstract

Deep learning (DL) requires labeled data. Labeling medical images requires medical expertise, which is often a bottleneck. It is therefore useful to prioritize labeling those images that are most likely to improve a model's performance, a practice known as instance selection. Here we introduce ENRICH, a method that selects images for labeling based on how much novelty each image adds to the growing training set. In our implementation, we use cosine similarity between autoencoder embeddings to measure that novelty. We show that ENRICH achieves nearly maximal performance on classification and segmentation tasks using only a fraction of available images, and outperforms the default practice of selecting images at random. We also present evidence that instance selection may perform categorically better on medical vs. non-medical imaging tasks. In conclusion, ENRICH is a simple, computationally efficient method for prioritizing images for expert labeling for DL.

Domain-guided data augmentation for deep learning on medical imaging

“…Using our pipeline ( Methods , Fig 1a ), we created new hybrid images that were realistic overall with only small cut-paste combination artifacts ( Fig 1b ). Adjacent frames in an ultrasound video from the same patient will generally have very similar heart structure [ 21 ], so we randomly chose only one frame per view per patient ID for target views and five frames per patient ID for NT view to generate a set of candidate images; these frames also had to pass quality control ( Methods ). Thus, we used only five percent of the original training data to synthesize thousands of new images such that the new hybrid training dataset was approximately the size of the original training dataset ( Table 1 ).…”

“…Efficient use of training data in model training can lighten data labeling burden, especially when combined with other strategies for training dataset curation [ 21 ]. This is particularly advantageous in the medical domain where there is a scarcity of data and experts to label this data.…”

Deep learning model for prenatal congenital heart disease (CHD) screening generalizes to the community setting and outperforms clinical detection

“…This study has several strengths. First, the diversity of the imaging cohort(23) represents a robust test for the DL model. This cohort is external to the dataset the model was trained on, differs with respect to image formats and scanning protocol, and represents imaging from several clinics and sonographers.…”