Arikan, Mustafa; (2025) Deep Active Learning Enhanced Robust Bioimage Analysis. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Arikan_Thesis.pdf - Submitted Version Download (19MB) | Preview

Abstract

This thesis addresses two critical challenges in bioimage analysis: data efficiency and robustness, focusing on deep learning applications for optical coherence tomography (OCT) and bioimaging in general. The adoption of deep learning in this field has been hindered by the scarcity of large, annotated datasets and the unique characteristics of images, such as variations in quality, noise levels, and data types. These factors complicate both data collection and model generalization across diverse imaging conditions and patient populations. To overcome these barriers, this research develops a comprehensive workflow for data management, annotation, curation, and model training. The contributions of this work include: (1) Introducing an active learning strategy that incorporates human experts in the loop to enhance model training. (2) Creation of a framework to enable efficient image grading by domain experts. (3) Development of software tools and algorithms for data management and preparation, applicable to OCT devices like the Heidelberg Spectralis. (4) Curation of a large, annotated multi-disease and multi-grade dataset for OCT available for research use. We develop two similarity-based active learning approaches: SimUNet for segmentation and classification tasks, and SimDet for object detection. SimUNet employs image-based similarity combined with uncertainty, while SimDet uses embeddings from a pre-trained model. Both approaches prioritize diverse, representative samples while minimizing annotation redundancy. We validate these methods across multiple tasks including retinal layer segmentation, disease classification, and object detection for cell imaging. Our evaluation framework includes tests for adversarial robustness and performance on simulated datasets with increasing noise levels. Our experiments demonstrate significant improvements in both data efficiency and model robustness. SimUNet reduced the required training samples by 26.8-85.7% across different test sets while maintaining performance equivalent to random sampling. In classification tasks, SimUNet achieved the same performance with 35.7-82.4% fewer samples. For object detection in cell imaging, SimDet required 71-83% fewer samples compared to random selection. Furthermore, both approaches showed superior robustness on out-of-distribution datasets, with consistent performance across varying image quality. These results confirm that our similarity-based active learning approaches not only reduce annotation burden but also enhance model generalization to unseen data distributions. These findings have the potential to enhance the effectiveness of imaging studies across healthcare and research applications, making significant contributions to the field of biomedical image analysis.

Download activity - last month

Export as JSONCSVXML

Download activity - last 12 months

Export as JSONCSVXML

Downloads by country - last 12 months

Export as XMLCSVJSON

Archive Staff Only

View Item