Deep ensemble model for segmenting microscopy images in the presence of limited labeled data

Obtaining large amounts of high quality labeled microscopy data is expensive and time-consuming. To overcome this issue, we propose a deep ensemble model which aims to utilise limited labeled training data. We train multiple identical Convolutional Neural Network (CNN) segmentation models on training data that is partitioned into folds in two steps. First, the data is split based on sample diversity or expert knowledge reflecting the possible {\em modes} of the underlying data distribution. In the second step, these partitions are split into random folds like in a cross-validation setting. Segmentation models based on the U-net architecture are trained on each of these resulting folds yielding the candidate models for our deep ensemble model which are aggregated to obtain the final prediction. The proposed deep ensemble model is compared to relevant baselines, in their ability to segment interneurons in microscopic images of mice spinal cord, showing improved performance on an independent test set.