CAMEL: Confidence-Aware Multi-task Ensemble Learning with Spatial Information for Retina OCT Image Classification and Segmentation

Juho Jung, Migyeong Yang, Hyunseon Won, Jiwon Kim, Jinyoung Han

Abstract:

Precise retina Optical Coherence Tomography (OCT)
image classification and segmentation are important for di-
agnosing various retinal diseases and identifying specific
regions. Alongside comprehensive lesion identification, re-
ducing the predictive uncertainty of models is crucial for
improving reliability in clinical retinal practice. However,
existing methods have primarily focused on a limited set
of regions identified in OCT images and have often faced
challenges due to aleatoric and epistemic uncertainty. To
address these issues, we propose CAMEL (Confidence-
Aware Multi-task Ensemble Learning), a novel frame-
work designed to reduce task-specific uncertainty in multi-
task learning. CAMEL achieves this by estimating model
confidence at both pixel and image levels and leveraging
confidence-aware ensemble learning to minimize the un-
certainty inherent in single-model predictions. CAMEL
demonstrates state-of-the-art performance on a compre-
hensive retinal OCT image dataset containing annotations
for nine distinct retinal regions and nine retinal diseases.
Furthermore, extensive experiments highlight the clini-
cal utility of CAMEL, especially in scenarios with mini-
mal regions, significant class imbalances, and diverse re-
gions and diseases. Our code is publicly available at:
https://github.com/DSAIL-SKKU/CAMEL.