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).
The Challenge of Sparse Labeling in FIB-SEM
In biological imaging, specifically Focused Ion Beam Scanning Electron Microscopy (FIB-SEM), segmenting subcellular structures is a significant bottleneck. Researchers typically rely on two primary training formulations, each with distinct trade-offs:
- One-hot Encoding (Classification): The model predicts class membership for each pixel. This approach is highly flexible; knowing a pixel belongs to a mitochondrion implies it is not a nucleus. This allows models to learn from "sparse" data—where only specific objects are labeled—by treating unlabeled areas as negative examples for the target class.
- Signed Boundary Distances (Regression): The model predicts the distance of a pixel from an object's boundary. This is superior for instance segmentation via "watershed" post-processing, as it helps separate touching objects. However, calculating these distances requires dense labels; if a dataset is only partially annotated, it cannot easily be used for distance-based training.
Methodology: The Hot-Distance Approach
To overcome the "data bottleneck" where partially labeled data is discarded in distance-based pipelines, we propose Hot-Distance, a multi-task learning strategy that performs both classification and distance regression simultaneously.
Shared Neural Backbone
The architecture utilizes a shared neural network backbone for both tasks. By sharing parameters, the features learned from sparse "negative" examples in the one-hot branch directly inform the internal representations used by the distance branch. The network only diverges into two separate "heads" at the final layer:
- Classification Head: Outputs class probabilities (one-hot).
- Distance Head: Outputs signed distances bounded by a hyperbolic tangent (tanh) function.
Leveraging Mutual Exclusivity
Because biological structures (e.g., nuclei and mitochondria) are mutually exclusive, the one-hot branch can be trained on "non-target" crops. If a training image only contains labeled mitochondria, the model still learns that those pixels have a 0% probability of being a nucleus, effectively utilizing "negative knowledge" to refine the model's spatial understanding.
Comparative Analysis
Hot-Distance effectively bridges the gap between data efficiency and post-processing utility. By breaking the requirement for fully annotated fields of view, the backbone can learn from a much larger and more diverse set of partially labeled data while maintaining the mathematical smoothness required for watershed segmentation.
Conclusion and Future Work
Hot-Distance provides a conceptual framework for maximizing the utility of existing, sparsely labeled biological datasets. While this initial report establishes the architectural proof-of-concept, ongoing empirical studies are focused on quantifying performance gains (e.g., Dice scores and Variation of Information) across various public benchmarks like COSEM. Future iterations may explore integrating additional loss functions to further refine boundary precision in complex subcellular environments.
Summary updated