Boosting Overlapping Organoid Instance Segmentation Using Pseudo-Label Unmixing and Synthesis-Assisted Learning

TL;DR

The paper addresses the challenge of overlapping organoid instance segmentation under limited labeled data by introducing Pseudo-Label Unmixing (PLU) to detect and decompose erroneous pseudo-labels and employing contour-based image synthesis to preserve overlap structure. It adapts Synthesis-Assisted Semi-Supervised Learning (SA-SSL) to organoids, incorporating instance-level augmentations and a pseudo-label guided synthesis pipeline to generate high-quality synthetic data. Through two curated datasets, OrganoSegment and M-OrgaQuant, the authors demonstrate that their SA-SSL framework achieves state-of-the-art results, rivaling fully supervised models while using only a fraction of labeled data; ablations verify the effectiveness of PLU, contour-based synthesis, and augmentation-aware training. The work advances scalable, label-efficient analysis of organoids, with implications for high-throughput drug screening and precision medicine, and it provides a rigorous evaluation of distribution alignment via FID and related metrics.

Abstract

Organoids, sophisticated in vitro models of human tissues, are crucial for medical research due to their ability to simulate organ functions and assess drug responses accurately. Accurate organoid instance segmentation is critical for quantifying their dynamic behaviors, yet remains profoundly limited by high-quality annotated datasets and pervasive overlap in microscopy imaging. While semi-supervised learning (SSL) offers a solution to alleviate reliance on scarce labeled data, conventional SSL frameworks suffer from biases induced by noisy pseudo-labels, particularly in overlapping regions. Synthesis-assisted SSL (SA-SSL) has been proposed for mitigating training biases in semi-supervised semantic segmentation. We present the first adaptation of SA-SSL to organoid instance segmentation and reveal that SA-SSL struggles to disentangle intertwined organoids, often misrepresenting overlapping instances as a single entity. To overcome this, we propose Pseudo-Label Unmixing (PLU), which identifies erroneous pseudo-labels for overlapping instances and then regenerates organoid labels through instance decomposition. For image synthesis, we apply a contour-based approach to synthesize organoid instances efficiently, particularly for overlapping cases. Instance-level augmentations (IA) on pseudo-labels before image synthesis further enhances the effect of synthetic data (SD). Rigorous experiments on two organoid datasets demonstrate our method's effectiveness, achieving performance comparable to fully supervised models using only 10% labeled data, and state-of-the-art results. Ablation studies validate the contributions of PLU, contour-based synthesis, and augmentation-aware training. By addressing overlap at both pseudo-label and synthesis levels, our work advances scalable, label-efficient organoid analysis, unlocking new potential for high-throughput applications in precision medicine.

Figures (7)

• Figure 1: Labels and pseudo labels of organoids. Incorrect pseudo labels of overlapping structures are highlighted by red rectangles.
• Figure 2: Illustration of pseudo-label correction. An overlapping judgment branch detects erroneous pseudo-labels, which are subsequently rectified through a decomposition branch to obtain a corrected version. FC: Fully connected, $p_{i}$: Model-predicted overlap judgments, $y_{i}$: Ground-truth overlap judgments.
• Figure 3: Workflow of our pseudo-label guided Image synthesis. Labeled data is categorized into four groups based on organoid transparency and focus state. Contour images of categorized organoids are generated and integrated into the generative adversarial training of the synthesis model, along with corresponding images. The modified model uses augmented images from corrected pseudo-labels to generate images. Here, instance masks are color-coded to distinguish individual organoids, while contour images use separate color schemes to represent different categories. DS: Down-sampling, IA: Instance-level Augmentations.
• Figure 4: Framework of SA-SSL. A teacher model (T) is initially trained on a limited set of labeled images. The trained teacher model generates pseudo-labels for unlabeled images. After the PLU process, new images are synthesized through pseudo-label guided image synthesis to augment the training dataset. A student model (S) is trained on a combination of real data, pseudo-labeled data, and synthetic data. The weights of the trained S are periodically updated to T via EMA.
• Figure 5: Changes in the evaluation metric scores ($\Delta S$) of the instance segmentation model after applying various data augmentation techniques under various labeled data conditions.