Confidence-Guided Semi-supervised Learning for Generalized Lesion Localization in X-Ray Images

Abhijit Das, Vandan Gorade, Komal Kumar, Snehashis Chakraborty, Dwarikanath Mahapatra, Sudipta Roy

Abstract: In recent years pseudo label (PL) based semi-supervised (SS) methods have been proposed for disease localization in medical images for tasks with limited labeled data. However these models are not curated for chest x-rays containing anomalies of different shapes and sizes. As a result, existing methods suffer from biased attentiveness towards minor class and PL inconsistency. Soft labeling based methods filters out PLs with higher uncertainty but leads to loss of fine-grained features of minor articulates, resulting in sparse prediction. To address these challenges we propose AnoMed, an uncertainty aware SS framework with novel scale-invariant bottleneck (SIB) and confidence guided pseudo-label optimizer (PLO). SIB leverages base feature (\(\mathcal {F}_b\)) obtained from any encoder to capture multi-granular anatomical structures and underlying representations. On top of that, PLO refines hesitant PLs and guides them separately for unsupervised loss, reducing inconsistency. Our extensive experiments on cardiac datasets and out-of-distribution (OOD) fine-tuning demonstrate that AnoMed outperforms other state-of-the-art (SOTA) methods like Efficient Teacher and Mean Teacher with improvement of 4.9 and 5.9 in \(AP_{50:95}\) on VinDr-CXR data. Code for our architecture is available at https://github.com/aj-das-research/AnoMed.