Abstract: A common assumption in machine vision is that the training and test samples are drawn from the same distribution. However, there are many problems when this assumption is grossly violated, as in bio-medical applications where different acquisitions can generate drastic variations in the appearance of the data due to changing experimental conditions. This problem is accentuated with 3D data, for which annotation is very time-consuming, limiting the amount of data that can be labeled in new acquisitions for training. In this paper we present a multi-task learning algorithm for domain adaptation based on boosting. Unlike previous approaches that learn task-specific decision boundaries, our method learns a single decision boundary in a shared feature space, common to all tasks. We use the boosting-trick to learn a non-linear mapping of the observations in each task, with no need for specific a-priori knowledge of its global analytical form. This yields a more parameter-free domain adaptation approach that successfully leverages learning on new tasks where labeled data is scarce. We evaluate our approach on two challenging bio-medical datasets and achieve a significant improvement over the state of the art.
The Challenge of Domain Shift
In machine learning, we typically assume that training and test data originate from the same distribution. In real-world bio-medical imaging, this assumption is frequently violated. Factors such as varying experimental conditions or different anatomical regions create a "domain shift," where a classifier trained on one dataset fails when applied to another.
This issue is particularly problematic for 3D image stacks, where manual labeling is labor-intensive. Domain Adaptation (DA) seeks to solve this by leveraging a large, labeled "source" dataset to improve performance on a sparsely labeled "target" dataset.
Methodology: The Boosting-Trick
The authors propose a multi-task learning approach that assumes the underlying classification problem is identical across domains—only the feature representations have been distorted by the acquisition process.
Core Innovations:
- Shared Decision Boundary: Instead of learning different boundaries for different tasks, the model learns a single linear boundary in a common latent feature space.
- Task-Specific Mappings: The model uses a "boosting-trick" to learn non-linear functions that map data from specific domains into the shared space.
- Scalable Gradient Boosting: Unlike kernel-based methods that scale poorly with large data, this approach uses gradient boosting to iteratively add weak learners (like regression trees) that learn task-specific feature shifts. This makes the method highly scalable for massive 3D microscopy datasets.
Experimental Validation
The researchers evaluated their approach on two difficult bio-medical tasks. One notable experiment involved Path Classification, where the model used 2D aerial road images as a source domain to classify 3D neural axons in the target domain.
They also tested the method on Mitochondria Segmentation, transferring knowledge between EM stacks of different brain regions (Striatum to Hippocampus).
Key Results and Insights
- Efficiency: In path classification, the method achieved near-peak performance (comparable to 1,000 labeled samples) using only 70 training samples in the target domain.
- Accuracy: On the mitochondria dataset, the method outperformed standard pooling and complex multi-task learning algorithms in segmentation accuracy (Jaccard Index).
- Speed: Training was significantly faster than competitors. For mitochondria tasks, the approach took 2–15 minutes, whereas leading alternatives required over 7 hours due to heavy cross-validation needs.
Conclusion
By using boosting to perform non-linear feature alignment, this framework allows a single classifier to work across diverse domains. While currently optimized for feature shifts, future work may extend this to handle complex geometric transformations and multi-modal data where initial feature spaces differ entirely.
Summary updated