NeuroADDA: Active Discriminative Domain Adaptation in Connectomic

2025-03-08

AI Summary

High annotation costs and poor model generalizability hinder cross-species connectomics in electron microscopy. Method: We propose a source-domain-agnostic active discriminative domain adaptation framework. First, we theoretically establish that the Maximum Mean Discrepancy (MMD) between neural image distributions quantifies cross-species transfer feasibility; empirically, we discover a strong correlation between feature-space domain distance and phylogenetic distance. Leveraging this insight, we design an MMD-driven optimal source domain selection mechanism, integrated with source-agnostic active learning and fine-tuning of pretrained backbone networks. Results: Evaluated on six cross-species connectomics datasets, our method achieves a 25–67% reduction in Variation of Information (VI) using only four annotated samples—substantially outperforming from-scratch training. It significantly enhances segmentation efficiency and generalization under stringent annotation budgets.

Technology Category

• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning
• Cognitive Modeling & Cognitive Systems: Neural Spike Coding
• Computer Vision: Medical and Biological Imaging

Application Category

• Search and Retrieval-Augmented AI: Vertical and domain-specific search
• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods

Abstract

Training segmentation models from scratch has been the standard approach for new electron microscopy connectomics datasets. However, leveraging pretrained models from existing datasets could improve efficiency and performance in constrained annotation budget. In this study, we investigate domain adaptation in connectomics by analyzing six major datasets spanning different organisms. We show that, Maximum Mean Discrepancy (MMD) between neuron image distributions serves as a reliable indicator of transferability, and identifies the optimal source domain for transfer learning. Building on this, we introduce NeuroADDA, a method that combines optimal domain selection with source-free active learning to effectively adapt pretrained backbones to a new dataset. NeuroADDA consistently outperforms training from scratch across diverse datasets and fine-tuning sample sizes, with the largest gain observed at $n=4$ samples with a 25-67% reduction in Variation of Information. Finally, we show that our analysis of distributional differences among neuron images from multiple species in a learned feature space reveals that these domain"distances"correlate with phylogenetic distance among those species.

Problem

• Improving segmentation model efficiency in connectomics using pretrained models.
• Identifying optimal source domains for transfer learning via MMD analysis.
• Developing NeuroADDA for effective domain adaptation and active learning.

Innovation

• Uses Maximum Mean Discrepancy for domain transferability
• Introduces NeuroADDA for optimal domain adaptation
• Combines active learning with pretrained models