Harnessing the Power of Bayesian Neural Networks for Annotator Consensus Refinement to Enhance Meibomian Gland Dysfunction Classification

We developed a novel approach to address label noise and low inter-annotator consensus in medical image classification. Applying a Bayesian neural network (BNN), we classified Meibomian gland dysfunction (MGD) using over 4,000 images from the OCULUS Keratograph 5M and MGD-1k datasets. The MGD-1k dataset, consisting of 1,000 annotated images with eyelid and Meibomian gland masks, was used to train segmentation models for OCULUS images. These segmentations were then leveraged to train a four-class classification model on more than 4,000 images, with approximately 3,000 images for training and 1,000 reserved for validation and testing.The dataset posed significant challenges due to noisy labels and low inter-annotator agreement, with each image receiving 3 to 6 potentially conflicting labels. Initially, we averaged the annotators' labels to train the model. However, as the BNN learned more from the data, we refined the labels by progressively selecting the most confident label from the annotators' provided set, ensuring no label was altered outside the range assigned by the annotators. This approach enabled the model to better capture the nuances in the data, leading to an improvement of more than 11% in validation accuracy.While Bayesian neural networks can handle uncertainty, they are susceptible to inconsistent labels. By combining these models with an iterative, fully automated label refinement, we overcame this challenge, which led to significant performance improvement. To our knowledge, this is the first use of BNNs for noisy label refinement in this domain.