Instance-Aware Pseudo-Labeling and Class-Focused Contrastive Learning for Weakly Supervised Domain Adaptive Segmentation of Electron Microscopy

TL;DR

This paper addresses the challenge of segmenting numerous mitochondria instances across EM image domains with minimal annotation. It introduces a multitask WDA framework that jointly performs segmentation and center-point detection, using cross-task teaching and an instance-aware pseudo-labeling (IPL) strategy to curate diverse, reliable pseudo-labels, complemented by class-focused contrastive learning with foreground prototypes. The approach significantly narrows the gap to fully supervised performance, outperforming state-of-the-art UDA and WDA methods using only 15% sparse center-point annotations and showing robustness to annotation noise. Practically, this enables accurate, scalable mitochondria analysis across heterogeneous EM datasets without extensive target-domain labeling or adversarial training.

Abstract

Annotation-efficient segmentation of the numerous mitochondria instances from various electron microscopy (EM) images is highly valuable for biological and neuroscience research. Although unsupervised domain adaptation (UDA) methods can help mitigate domain shifts and reduce the high costs of annotating each domain, they typically have relatively low performance in practical applications. Thus, we investigate weakly supervised domain adaptation (WDA) that utilizes additional sparse point labels on the target domain, which require minimal annotation effort and minimal expert knowledge. To take full use of the incomplete and imprecise point annotations, we introduce a multitask learning framework that jointly conducts segmentation and center detection with a novel cross-teaching mechanism and class-focused cross-domain contrastive learning. While leveraging unlabeled image regions is essential, we introduce segmentation self-training with a novel instance-aware pseudo-label (IPL) selection strategy. Unlike existing methods that typically rely on pixel-wise pseudo-label filtering, the IPL semantically selects reliable and diverse pseudo-labels with the help of the detection task. Comprehensive validations and comparisons on challenging datasets demonstrate that our method outperforms existing UDA and WDA methods, significantly narrowing the performance gap with the supervised upper bound. Furthermore, under the UDA setting, our method also achieves substantial improvements over other UDA techniques.

Figures (5)

• Figure 1: Weakly-Supervised Domain Adaptation (WDA) for mitochondria segmentation in EM images. Sparse center points on a few object instances are utilized as the weak labels on target training data.
• Figure 2: Overview of the proposed method for weakly-supervised cross-domain adaptation. Under a multitask learning framework, an auxiliary center detection task is utilized to achieve instance-aware pseudo-label selection for the self-training on the segmentation head. class-focused conservative learning through pixel-to-prototype alignment is introduced to conduct feature-level domain alignment and improve the compactness of category feature representations.
• Figure 3: Qualitative comparison results. Green: true positives; Red: false positives; White: false negatives.
• Figure 4: Visual comparison of the ablated versions of our proposed approach.
• Figure 5: Robustness to location deviations of the sparse point annotation. The performance of our model (15%) on EPFL Hippocampus$\rightarrow$MitoEM-R is evaluated when randomly displacing the sparse point annotations by $d$ pixels from their respective centers.