[Paper Review] Prefer-DAS: Learning from Local Preferences and Sparse Prompts for Domain Adaptive Segmentation of Electron Microscopy

Jiabao Chen, Shan Xiong+1 more|arXiv (Cornell University)|Feb 23, 2026

Advanced Electron Microscopy Techniques and Applications0 citations

TL;DR

Prefer-DAS introduces a promptable, preference-guided domain-adaptive segmentation framework for EM images that uses sparse point prompts, local/global human preferences, and self-learned supervision to improve cross-domain mitochondria segmentation, achieving near-supervised performance.

ABSTRACT

Domain adaptive segmentation (DAS) is a promising paradigm for delineating intracellular structures from various large-scale electron microscopy (EM) without incurring extensive annotated data in each domain. However, the prevalent unsupervised domain adaptation (UDA) strategies often demonstrate limited and biased performance, which hinders their practical applications. In this study, we explore sparse points and local human preferences as weak labels in the target domain, thereby presenting a more realistic yet annotation-efficient setting. Specifically, we develop Prefer-DAS, which pioneers sparse promptable learning and local preference alignment. The Prefer-DAS is a promptable multitask model that integrates self-training and prompt-guided contrastive learning. Unlike SAM-like methods, the Prefer-DAS allows for the use of full, partial, and even no point prompts during both training and inference stages and thus enables interactive segmentation. Instead of using image-level human preference alignment for segmentation, we introduce Local direct Preference Optimization (LPO) and sparse LPO (SLPO), plug-and-play solutions for alignment with spatially varying human feedback or sparse feedback. To address potential missing feedback, we also introduce Unsupervised Preference Optimization (UPO), which leverages self-learned preferences. As a result, the Prefer-DAS model can effectively perform both weakly-supervised and unsupervised DAS, depending on the availability of points and human preferences. Comprehensive experiments on four challenging DAS tasks demonstrate that our model outperforms SAM-like methods as well as unsupervised and weakly-supervised DAS methods in both automatic and interactive segmentation modes, highlighting strong generalizability and flexibility. Additionally, the performance of our model is very close to or even exceeds that of supervised models.

Motivation & Objective

• Motivate accurate mitochondria segmentation across diverse EM domains with limited target annotations.
• Propose a flexible, promptable multitask model that supports automatic and interactive segmentation.
• Introduce local and sparse local preference learning to align model outputs with human judgments.
• Develop unsupervised and self-learning preference mechanisms to handle missing human feedback.
• Demonstrate strong performance against UDA, WDA, and SAM-like methods across multiple DAS benchmarks.

Proposed method

• Propose Prefer-DAS, a promptable multitask model with an image encoder, a point prompt encoder, a multitask decoder, and a segmentation head plus a center-point detection head.
• Use pseudo-prompt learning and mean-teacher self-training to leverage labeled source data and unlabeled target data.
• Incorporate prompt-guided contrastive learning to improve discriminative feature representations for prompts.
• Introduce Local direct Preference Optimization (LPO) and Sparse LPO (SLPO) to align segmentation with spatially varying human feedback.
• Add Unsupervised Preference Optimization (UPO) to learn from self-generated preferences when human feedback is unavailable.
• Enable both UDA and WDA modes, plus interactive segmentation with full/partial/no point prompts during inference.

Experimental results

Research questions

• RQ1 How can local and sparse local human preferences improve domain-adaptive segmentation of EM mitochondria?
• RQ2 Can a promptable model utilizing sparse points and preference learning achieve performance close to supervised models in cross-domain EM segmentation?
• RQ3 How effective are LPO, SLPO, and UPO in mitigating reward misspecification and improving segmentation under domain shift?
• RQ4 Does integrating a prompt-guided contrastive objective enhance discriminability of segmentation embeddings under weak supervision?

Key findings

• Prefer-DAS outperforms SAM-like methods and prior unsupervised/weakly-supervised DAS methods in automatic and interactive segmentation modes.
• The model achieves performance close to or exceeding supervised models on EM DAS benchmarks.
• Local and sparse local preferences, along with promptable learning, provide effective guidance under domain shifts and annotation budgets.
• LPO/SLPO successfully align segmentation with spatially varying human feedback with reduced labeling effort.
• UPO enables unsupervised preference optimization when human feedback is unavailable, maintaining strong performance.
• The framework supports both UDA and WDA, and allows interactive segmentation with variable prompting at inference.