Decomposition Sampling for Efficient Region Annotations in Active Learning

TL;DR

DECOMP addresses the high cost of dense region annotations in active learning by decomposing unlabeled images into class-specific components and guiding both image- and region-level sampling with class-wise confidence. The method achieves superior annotation efficiency across ROI classification, 2-D, and 3-D segmentation tasks, requiring far fewer annotations to reach near full-annotation performance compared with strong baselines. Its single forward pass per image, memory efficiency, and explicit handling of minority classes make it practical for medical imaging and other dense-prediction domains. The work is corroborated by experiments on BRACS, Cityscapes, and KiTS23, with supplementary analyses confirming robustness to budgets and hyperparameters and highlighting improvements in minority-class coverage.

Abstract

Active learning improves annotation efficiency by selecting the most informative samples for annotation and model training. While most prior work has focused on selecting informative images for classification tasks, we investigate the more challenging setting of dense prediction, where annotations are more costly and time-intensive, especially in medical imaging. Region-level annotation has been shown to be more efficient than image-level annotation for these tasks. However, existing methods for representative annotation region selection suffer from high computational and memory costs, irrelevant region choices, and heavy reliance on uncertainty sampling. We propose decomposition sampling (DECOMP), a new active learning sampling strategy that addresses these limitations. It enhances annotation diversity by decomposing images into class-specific components using pseudo-labels and sampling regions from each class. Class-wise predictive confidence further guides the sampling process, ensuring that difficult classes receive additional annotations. Across ROI classification, 2-D segmentation, and 3-D segmentation, DECOMP consistently surpasses baseline methods by better sampling minority-class regions and boosting performance on these challenging classes. Code is in https://github.com/JingnaQiu/DECOMP.git.

Figures (11)

• Figure 1: (a) al builds the annotation set and model iteratively. The last trained model is used to select new annotation samples, which are then labeled and added to the existing annotations for a new round of model update. (b) DECOMP enhances annotation diversity by querying annotations across different classes. It first segments an unlabeled image $X_i$ into class components using predictions from the model trained in the previous al cycle $h^{(t-1)}$. Class-wise sampling weights $\mathbf{w}$ are computed based on the proportion of high-confidence class predictions to prioritize annotations for uncertain classes. To select the most informative image, a score $s_i$ is determined for each unlabeled image by summing its predicted class frequencies, weighted by the sampling weights. For each selected image, regions are sequentially identified by first sampling a class based on $\mathbf{w}$ and then selecting the region that best represents it.
• Figure 2: Results on BRACS test set across $120, 90, 30, 20$ cycles for per-cycle annotation budget combinations $n_\text{image}\in\{1, 1, 5, 5\}$ and $n_\text{region}\in\{15, 15, 30, 30\}$, respectively. Weighted F1 as a function of processed WSIs (%) for different methods. Scatter points are denser in later cycles because all $68$ validation wsi have been fully annotated, leaving only training wsi in $\mathcal{U}$. Means over five runs.
• Figure 3: Results on Cityscapes validation set across $40$ and $25$ cycles for al budget combinations $n_\text{image}\in\{25, 50\}$ and $n_\text{region}=10$, respectively. miou as a function of annotated area (%) for different methods. Means over three runs.
• Figure 4: Results on KiTS23 across $45$ and $30$ cycles for per-cycle annotation budget combinations $n_\text{image}\in\{10, 20\}$ and $n_\text{region}=10$, respectively. DSC as a function of annotated area (%) for different methods. Means over five-fold cross validation.
• Figure 5: Left: Effect of DECOMP image selection, replacing uncertainty image selection in UNCERT, DIVERS(cluster), and DIVERS(core-set) consistently improves performance (compare solid–dashed pairs). Right: Effect of DECOMP region selection, replacing UNCERT’s region selector with DECOMP's yields performance gains (UNCERT_DECOMP vs. UNCERT). (Cityscapes)