Uncertainty-driven active learning in a deep semi-supervised framework for WCE image classification
Wireless Capsule Endoscopy (WCE) image classification using deep learning models is hindered by data scarcity and model uncertainty. Labelling medical images is costly and time-consuming, limiting the availability of labelled data for training. To address these challenges, this work proposes ACT-WIS...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-09-01
|
| Series: | Results in Engineering |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2590123025022467 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Wireless Capsule Endoscopy (WCE) image classification using deep learning models is hindered by data scarcity and model uncertainty. Labelling medical images is costly and time-consuming, limiting the availability of labelled data for training. To address these challenges, this work proposes ACT-WISE, an active inter-consistency-driven semi-supervised learning framework that integrates Active Learning (AL) with Deep Learning (DL). ACT-WISE uses a teacher-student training methodology in which the model improves consistency by learning structural and semantic correlations from perturbations of unlabelled images. To reduce model uncertainty, batch acquisition selects the most informative samples from an unlabelled data pool based on minimum redundancy and maximum predictive entropy. Unlike traditional semi-supervised methods, ACT-WISE dynamically refines its selection strategy, optimising both label efficiency and model reliability. Based on the experimental results on the Kvasir-Capsule dataset for WCE image classification, the proposed ACT-WISE model demonstrates superior performance by achieving a classification accuracy of 0.97 and an AUC of 0.95, outperforming prior approaches. In addition, this work uses Monte Carlo dropout for model uncertainty estimation and assesses calibration reliability using Expected Calibration Error (ECE), providing interpretable and trustworthy prediction confidence scores. These results demonstrate that, with little manual annotation, ACT-WISE can greatly improve automatic anomaly detection in capsule endoscopy imaging, hence enhancing diagnostic support. |
|---|---|
| ISSN: | 2590-1230 |