Hot-Distance: Combining One-Hot and Signed Distance Embeddings for Segmentation

TL;DR

The paper addresses the data bottleneck in segmentation tasks, particularly for subcellular structures in FIB-SEM, where dense annotations are expensive. It proposes Hot-Distance, a multi-task target that shares a backbone to predict both one-hot encodings and signed boundary distance maps, with a final-layer split. This approach enables learning from sparse and non-target data while preserving the benefits of both targets, including gradient-driven learning and post-processing via watershed. The authors provide an implementation and outline plans for empirical comparisons against standard one-hot and signed-distance losses, signaling practical impact for building richer training datasets.

Abstract

Machine learning models are only as good as the data to which they are fit. As such, it is always preferable to use as much data as possible in training models. What data can be used for fitting a model depends a lot on the formulation of the task. We introduce Hot-Distance, a novel segmentation target that incorporates the strength of signed boundary distance prediction with the flexibility of one-hot encoding, to increase the amount of usable training data for segmentation of subcellular structures in focused ion beam scanning electron microscopy (FIB-SEM).

Figures (1)

• Figure 1: Advantages of network prediction targets. (a) One-hot embeddings can represent true negative examples, even in the absence of known positive examples of a class. Here, true negatives are known for the nucleus class (red striped label) based on the true positive examples of the mutually exclusive mitochondria class (blue label). (b) Signed tanh boundary distances present smooth gradients for network predictions that can allow for produce instance segmentations via watershed post-processing.