UD-MIL: Uncertainty-Driven Deep Multiple Instance Learning for OCT Image Classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Ophthalmology

Background:

• Expert annotation of Optical Coherence Tomography (OCT) images for deep learning is costly and time-consuming.
• Leveraging volume-level labels for OCT image classification remains a challenge.

Purpose of the Study:

• To propose a weakly supervised deep learning framework with uncertainty estimation for macula-related disease classification using only volume-level OCT labels.
• To develop a robust instance classifier capable of detecting abnormal instances and extracting representative features.

Main Methods:

• A convolutional neural network (CNN) based instance-level classifier refined using an uncertainty-driven deep multiple instance learning (MIL) scheme.
• Integration of uncertainty evaluation into MIL for robust instance classification.
• A recurrent neural network (RNN) for bag-level prediction using instance features, considering local and global representations.

Main Results:

• Achieved high volume-level accuracy (95.1%), F1-score (0.939), and AUC (0.990) on the Heidelberg-DME dataset.
• Attained comparable results on the Triton-DME dataset (95.1% accuracy, 0.935 F1-score, 0.986 AUC).
• Demonstrated competitive performance on a public age-related macular degeneration OCT dataset.

Conclusions:

• The proposed weakly supervised deep learning framework effectively classifies macula-related diseases from OCT images using only volume-level labels.
• The method's ability to incorporate uncertainty estimation enhances the robustness of instance classification within MIL.
• The framework shows significant potential as an efficient screening tool in clinical practice for ophthalmological conditions.