Uncertainty-based Deep Active Learning for Retinal Layer Segmentation

Abstract

Purpose : Progress of deep learning (DL) research in medicine is often hampered by the time required to generate the large amounts of labelled data for model training. To address this, we propose an active learning method to optimize training efficiencies while minimizing labelling efforts. We employed model uncertainty as an objective metric for prioritizing the candidate training images. Intuitively, images giving rise to higher levels of model uncertainty during prediction would contain more informative features that could help accelerate the rate at which a DL model can learn. We evidence the benefit of active learning on retinal-layer segmentation in optical coherence tomography (OCT) images.

Methods : 20 AMD, 20 DME and 20 normal OCT volumes (16-61 scans/volume) were obtained using the Heidelburg Spectralis device. We used a 12-12-36 split over the volumes for training, validation and testing. We quantified gains from active learning by comparing segmentation performance (intersection-over-union (IoU)) of a DL model trained with images selected by high uncertainty against a baseline where training samples were selected at random. Starting with a DL model trained on a prior subset of training images, we employed the Bayesian Dropout method to estimate the trained model’s uncertainty when deployed on the remaining candidates of unseen training images. We select an increment of additional images based on high uncertainty values for model retraining. In the baseline experiment, additional images were selected at random. We continue to iterate the process and record the test performance over 4 increments and over 3 experimental repeats.

Results : Active learning produce higher IoU at the lower training sets sizes. This trend was greatest for DME scans.

Conclusions : In general, active learning is a principled method that can help minimize the burden of expensive and time-consuming data labelling and, in the process, can help accelerate deep learning research in medicine. In our results, while active learning provides trends of efficiency, more experimental repeats are needed to determine significance.

This is a 2021 ARVO Annual Meeting abstract.