Finally Outshining the Random Baseline: A Simple and Effective Solution for Active Learning in 3D Biomedical Imaging

TL;DR

This work tackles the high annotation cost of 3D biomedical image segmentation by proposing ClaSP PE, a simple active-learning query strategy that combines class-stratified sampling and log-scale power noising with an exponential schedule. By selecting patches across underrepresented classes and encouraging early query diversity, ClaSP PE consistently outperforms improved random baselines on the nnActive benchmark while maintaining annotation efficiency. The authors validate generalization through a roll-out study on unseen datasets using predefined deployment guidelines, demonstrating robust, near-production-ready performance without dataset-specific tuning. The open-source implementation and practical guidelines aim to make ClaSP PE a strong, accessible baseline for future AL research and real-world deployment in 3D biomedical imaging.

Abstract

Active learning (AL) has the potential to drastically reduce annotation costs in 3D biomedical image segmentation, where expert labeling of volumetric data is both time-consuming and expensive. Yet, existing AL methods are unable to consistently outperform improved random sampling baselines adapted to 3D data, leaving the field without a reliable solution. We introduce Class-stratified Scheduled Power Predictive Entropy (ClaSP PE), a simple and effective query strategy that addresses two key limitations of standard uncertainty-based AL methods: class imbalance and redundancy in early selections. ClaSP PE combines class-stratified querying to ensure coverage of underrepresented structures and log-scale power noising with a decaying schedule to enforce query diversity in early-stage AL and encourage exploitation later. In our evaluation on 24 experimental settings using four 3D biomedical datasets within the comprehensive nnActive benchmark, ClaSP PE is the only method that generally outperforms improved random baselines in terms of both segmentation quality with statistically significant gains, whilst remaining annotation efficient. Furthermore, we explicitly simulate the real-world application by testing our method on four previously unseen datasets without manual adaptation, where all experiment parameters are set according to predefined guidelines. The results confirm that ClaSP PE robustly generalizes to novel tasks without requiring dataset-specific tuning. Within the nnActive framework, we present compelling evidence that an AL method can consistently outperform random baselines adapted to 3D segmentation, in terms of both performance and annotation efficiency in a realistic, close-to-production scenario. Our open-source implementation and clear deployment guidelines make it readily applicable in practice. Code is at https://github.com/MIC-DKFZ/nnActive.

Figures (21)

• Figure 1: Overview of the ClaSP PE query strategy. We overcome two key limitations of standard uncertainty-based Active Learning methods (e.g. Predictive Entropy), class imbalance and low diversity of the queries, by adding two simple modifications: (1) class-stratified sampling for 66% of the query budget based on predicted class probabilities, and (2) a scheduler decreasing the noise for score perturbation via log-scale power noising to enhance diversity during query selection.
• Figure 2: ClaSP PE delivers substantial performance improvements without sacrificing annotation efficiency. The plots show average method rankings (lower is better) with standard error for AUBC, Final Dice, and FG-Eff across the nnActive benchmark. Results are aggregated over 4 datasets, 3 Label Regimes, and 2 query patch sizes, each evaluated with 4 random seeds, providing robust estimates of method performance. The brackets indicate groups of methods that do not differ significantly based on a post-hoc Nemenyi test at significance level $0.05$.
• Figure 3: ClaSP PE consistently outperforms both random and AL baselines across the nnActive benchmark. The Pairwise Penalty Matrix summarizes statistically significant wins and losses from pairwise t-tests (p=0.05) between methods. Results are aggregated over 24 distinct AL settings on the nnActive benchmark, including 4 datasets $\times$ 3 Label Regimes $\times$ 2 query patch sizes. Remaining lose scenarios against Random 66% FG stem from challenging Low-Label settings on the AMOS dataset (discussed in \ref{['sec:amos']}).
• Figure 4: Longer training amplifies the advantage of ClaSP PE over random selection. Shown are fractions of significant wins, losses, and resulting ties of ClaSP PE against improved random baselines on the AMOS dataset, as computed via the PPM. We compare models trained for 200 (left) and 500 (right) epochs, as well as different Label Regimes (color-coded). Each Label Regime carries 33% of the entire fraction of experiments which is then divided into wins, losses and ties. While at 200 epochs ClaSP PE loses on 60% of the experiments to FG66 and ties in the rest, it outperforms Random FG 66% in 20%, ties in 48% and loses in only 32% when trained for 500 epochs.
• Figure 5: ClaSP PE achieves the best trade-off between segmentation quality and annotation efficiency. Average method rankings on the nnActive Main benchmark (4 datasets $\times$ 3 Label Regimes $\times$ 1 query patch size), with additional method variants, Cla PE 66%, Cla PE 33% and ClaP PE.