Source-linked AI summary
Weakly supervised neural network: segmentation of complex structures in X-ray microCT
Daniele Rusconi, Michela Ascolese, Stephanie Fest-Santini, Alberto Bravin, Maurizio Santini
TL;DR
Dense pixel-level annotation makes segmentation of small, low-contrast glomeruli in microCT costly, motivating annotation-efficient learning. The study adapts a 2D nnU-Net to combine sparse dot annotations with limited fully segmented images. Mixed supervision improves segmentation toward the fully supervised baseline, while the authors identify loss design for sparse labels as an important limitation and opportunity.
Problem
Segmenting small, low-contrast glomeruli in microCT requires costly dense pixel-level annotations, creating a need for less annotation-intensive approaches.
Method
A 2D nnU-Net was trained using a larger weakly annotated dataset of sparse dots alongside a smaller fully segmented dataset.
Results
0.5748 ± 0.1358 Dice with 85% weak and 15% fully segmented images approached the 0.7710±0.0499 fully supervised baseline.
Takeaways & Limitations
Weak supervision provides a meaningful learning signal while a small number of high-quality annotations substantially improves accuracy and reduces annotation effort.
Takeaways & Limitations
Dot-based labels mismatch full-object Dice evaluation, and no systematic sweep was performed for the loss-mixing parameter λ.
Abstract
from arXiv · showhide
Segmentation of complex structures in X-ray tomographic data is a fundamental task in biomedical research, but it often requires large amounts of precisely annotated data, making fully supervised approaches costly and difficult to scale. In this study, weakly supervised deep learning is investigated as a strategy to reduce annotation effort while maintaining accurate segmentation. A two-dimensional convolutional neural network based on the nnU-Net framework was adapted to a weak supervision setting using sparse dot-based annotations, complemented by a limited number of fully segmented images. The approach was evaluated on high-resolution microCT slices of rat kidneys, targeting the segmentation of renal glomeruli, which are small, low-contrast anatomical structures. Results indicate that weak supervision provides a meaningful learning signal, enabling reliable localization of glomeruli even in the absence of dense labels. Incorporating a small set of high-quality annotations substantially improves segmentation performance, approaching that of a fully supervised model. These findings highlight the potential of weakly supervised learning as an annotation-efficient strategy for the analysis of complex structures in X-ray tomographic data, and suggest that alternative loss formulations tailored to sparse annotations may further enhance performance.
1 Introduction
Renal glomeruli are important targets for kidney research, but their small size and low contrast make microCT segmentation difficult and dense annotation costly. The study therefore investigates weakly supervised nnU-Net segmentation using sparse annotations.
- MicroCT provides high-resolution three-dimensional imaging, but glomeruli remain difficult to segment because they are small and low contrast.
- Fully supervised deep learning requires costly, time-consuming pixel-level annotations.
- Weak supervision uses inexpensive sparse or imprecise labels while retaining semantic information.
- The study adapts an nnU-Net-based framework to automate glomeruli segmentation in rat kidney microCT data.
- The work examines the trade-off between annotation effort and segmentation performance using sparse annotations instead of exhaustive pixel-wise labels.
2 Data acquisition
Rat kidney samples were imaged with high-resolution microCT, and preprocessing standardized slice intensities for subsequent analysis.
- Three male Sprague-Dawley rats were used, under institutional and national and international animal-care approvals.
- MicroCT acquisition used isotropic 6.81 × 6.81 × 6.81 𝜇m3 resolution with 3200 projections per tomography.
- Detector charge accumulation and spatial non-uniformity in detector sensitivity and X-ray source flux were corrected during acquisition.
- Slice intensity normalization reduced residual contrast-fluid effects and standardized pixel-value ranges.
3 Methods
The method combines a larger set of rapid dot annotations with a smaller expert-refined dataset to train a 2D nnU-Net under weak supervision. A mixed loss applies full Dice supervision to dense labels and localized supervision around sparse dots.
- Approximately 6000 images from three samples were analyzed, with annotated subsets used across supervision regimes.
- About 300 slices received expert-refined pixel-wise masks, requiring 15–20 minutes per slice on average.
- Approximately 800 slices were weakly annotated by non-experts in no more than 2 minutes per image using 8-pixel circular dots.
- The combined datasets enable weakly supervised training while reducing annotation burden and leveraging structural priors from high-quality labels.
- A 2D nnU-Net with an encoder–decoder architecture and skip connections was used for individual microCT slices.
- The weak Dice loss evaluates pixels near dot annotations, encouraging high predicted probabilities there while leaving unannotated regions unconstrained.
- The total loss balances fully and weakly annotated data with λ=0.5, selected empirically without a systematic hyperparameter sweep.
4 Results
Weak, mixed, and fully supervised models were compared using pseudo Dice. Mixed supervision substantially improved over weak-only training, although fully supervised training remained strongest.
- Three strategies were evaluated: dot-only weak supervision, mixed supervision, and fully supervised training.
- 0.3223 ± 0.0765 Dice was obtained with weak annotations alone, reflecting limited overlap with reference segmentations.
- 0.5748 ± 0.1358 Dice was achieved with 85% weak annotations and 15% fully segmented images.
- 0.7710±0.0499 Dice was achieved by the fully supervised model, providing the upper performance bound for this task.
- A limited set of fully annotated images significantly boosted performance, creating a trade-off between annotation effort and segmentation quality.
5 Conclusions
This study evaluates weakly supervised 2D nnU-Net segmentation of renal glomeruli using dot-based annotations, mixed supervision, and fully supervised training. Weak supervision provides a useful localization signal, while high-quality annotations improve accuracy; point-aligned losses may further help.
- 5 Conclusions: A 2D nnU-Net model was evaluated for renal glomerulus segmentation under weak, mixed, and fully supervised training regimes.The task uses rat kidney microCT images and compares dot-based annotations with fully annotated images.
- 5 Conclusions: Weak-only supervision identified regions likely to contain glomeruli without explicit localization constraints.This indicates that sparse dot annotations provide a meaningful signal for anatomical localization.
- 5 Conclusions: Dice loss may be suboptimal for sparse point annotations because it evaluates full-object pixel overlap against incomplete labels.Centroid- or distance-based losses are proposed as alternatives better aligned with point annotations.
- 5 Conclusions: Adding a small number of high-quality annotations substantially improved accuracy, approaching the fully supervised baseline.The conclusion links mixed supervision to improved segmentation performance while reducing reliance on exhaustive labels.