WaveFuse-AL: Cyclical and Performance-Adaptive Multi-Strategy Active Learning for Medical Images

TL;DR

WaveFuse-AL tackles the inconsistency of single-query active-learning strategies in medical imaging by fusing multiple strategies through a cyclical, sinusoidally-prioritized weighting scheme combined with performance-driven adaptation. The approach blends BALD, BADGE, Entropy, and CoreSet with phase-shifted priors and smoothed performance traces to smoothly allocate influence across rounds, expressed as $w_s(t)$ and batch quotas $q_s$. Evaluation across APTOS-2019, RSNA Pneumonia Detection, and ISIC-2018 shows consistent gains over single-strategy and alternating baselines, with significant improvements on a majority of metrics at 10–20% labeling budgets, demonstrating improved sample efficiency. The method reduces annotation costs while maintaining or improving diagnostic and segmentation performance, suggesting practical impact for scalable, high-quality medical imaging pipelines.

Abstract

Active learning reduces annotation costs in medical imaging by strategically selecting the most informative samples for labeling. However, individual acquisition strategies often exhibit inconsistent behavior across different stages of the active learning cycle. We propose Cyclical and Performance-Adaptive Multi-Strategy Active Learning (WaveFuse-AL), a novel framework that adaptively fuses multiple established acquisition strategies-BALD, BADGE, Entropy, and CoreSet throughout the learning process. WaveFuse-AL integrates cyclical (sinusoidal) temporal priors with performance-driven adaptation to dynamically adjust strategy importance over time. We evaluate WaveFuse-AL on three medical imaging benchmarks: APTOS-2019 (multi-class classification), RSNA Pneumonia Detection (binary classification), and ISIC-2018 (skin lesion segmentation). Experimental results demonstrate that WaveFuse-AL consistently outperforms both single-strategy and alternating-strategy baselines, achieving statistically significant performance improvements (on ten out of twelve metric measurements) while maximizing the utility of limited annotation budgets.

Figures (4)

• Figure 1: Overview of the proposed active learning framework for medical image analysis. In Phase-1, stratified splits are used to evaluate candidate models, selecting the best performing model. In Phase-2, sample selection is performed using a fusion of query strategies (Entropy, BALD, CoreSet, BADGE) combined via WaveFuse, producing acquisition scores for the selection of informative samples for labeling.
• Figure 2: Sinusoidal Strategy Weights across rounds
• Figure 3: t-SNE embeddings (top) and class-wise acquisition rates (bottom) comparing the proposed active-learning strategy with random sampling on APTOS-2019 (Class 0:No DR-49.3%, Class 1: Mild DR-10.1%, Class 2:Moderate DR- 27.3%, Class 3: Severe DR- 5.3%, and Class 4: Proliferative DR- 8.1%) and RSNA Pneumonia (Class 0: Normal - 67.73% and Class 1: Pneumonia - 32.27%). The labeled points are shown as colored solid circles, while unlabeled points are transparent circles with a solid border. Ours shows broader cluster coverage and reduced class bias.
• Figure 4: Performance comparison of active-learning strategies across three datasets (across all the rounds) where we select $~20\%$ of the total samples after 10 rounds. Our method consistently outperforms baselines across all datasets.