Exploiting data characteristics to improve automated optic nerve head segmentation and localization in OCT en face images

Abstract

Purpose : Machine learning (ML) literature more comprehensively focuses on improvements based on model architecture, optimization or learning strategies than on the underlying data. Domain knowledge for ML-based medical image analysis is of equal importance. We exploit data characteristics to improve optic nerve head (ONH) segmentation and localization.

Methods : In this work, instead of tuning the model architecture of a standard U-Net and related optimization and learning strategy, we focus on data engineering techniques to improve the model’s ONH segmentation and localization performance. We exploit the evidence-based area of possible ONH locations by cropping the input images from 256×256 pixels to the central 96×256 pixels region. To evaluate the performance of the data-centric approach, the binary cross entropy loss (CL) and the Tversky loss (TL) are utilized to optimize the model. The Dice's coefficient (DC) and the Euclidean distance (ED) between the predicted and true ONH centroid locations are calculated to evaluate the ONH segmentation performance and the localization performance of the model, respectively.

Results : The model was trained on 100, validated on 10 and tested on 10 en face projections from volumetric 60 degree widefield swept-source optical coherence tomography (SS-OCT) scans acquired with a 1.68 MHz prototype OCT device. The data sets were split at subject level. Compared to training on non-cropped data, with the data-centric approach, mean DC increased from 0.00 to 0.86 and from 0.85 to 0.91 for the CL and TL, respectively. Qualitative ONH segmentation results are shown in Fig. 1. The mean ED decreased from not-measurable (due to missing positive class predictions) to 2.70 pixels and from 8.79 to 1.27 pixels for the CL and TL, respectively. 2D distributions of ONH location errors are shown in Fig. 2.

Conclusions : Results demonstrate that the spatial restriction of the input images improves ONH segmentation and localization and is independent of the utilized loss functions. The results suggest that focusing not only on model architecture and optimization strategies but also on data engineering can improve the overall ML model performance.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

Figure 1. Qualitative ONH segmentation results on test set samples for a U-Net trained on the non-cropped full image sizes (a) or on cropped images (b).

 

Figure 2. Spatial distribution of ONH location errors.