Comparative analysis of out-of-distribution detection approaches in large scale retinal OCT datasets

Summarise with VIRAT
Authors: Víctor Bossa Muñoz
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Out-of-distribution (OOD) detection is essential for ensuring the safety and robustness of deep learning models in medical imaging. This thesis investigates the performance of four OOD detection methods—Cosine Distance, Mahalanobis Distance, Dirichlet-based uncertainty, and Entropy—on retinal OCT data. Experiments were conducted across three datasets (Kermany, OCTDL, OCTID) using ResNet-50, Vision Transformer (ViT), and the foundation model RETFound with an MLP classifier. Results show that OOD detection performance is highly dependent on both the model architecture and the chosen method. ViT combined with Cosine or Mahalanobis Distance consistently achieved the best results, particularly in far-OOD and blurred scenarios. Near- OOD detection proved more challenging due to significant overlap in feature distributions, especially in the OCTDL dataset. RETFound underperformed, probably due to the lack of fine-tuning and the simplicity of the MLP architecture. Visual analyses and t-SNE projections supported the quantitative findings, revealing strong separability for most ID and OOD samples, except in ambiguous cases. Mahalanobis Distance offered improved performance in complex scenarios due to its consideration of feature covariance, while Cosine Distance demonstrated the most stable behavior across datasets. Overall, this work highlights the importance of selecting well-calibrated models and detec- tion strategies tailored to specific data characteristics. Combining multiple OOD methods may further enhance reliability, particularly in clinical settings involving uncertain or novel inputs.

Fields of Study & Subjects

Adversarial Robustness in Machine Learning Retinal Imaging and Analysis Anomaly Detection Techniques and Applications

Publication details

Journal / source
University Library Linz repository (Johannes Kepler Universitat Linz)
Publication date
1 January 2026
Language
EN
Document type
Thesis