Deep learning-based automatic segmentation of stromal infiltrates and associated biomarkers on slit-lamp images of microbial keratitis

Abstract

Purpose : Accurate quantification of corneal ulcer biomarkers is expected to facilitate early and accurate disease classification and result in improved patient outcomes. To this end, we developed the first automatic deep learning based method to segment stromal infiltrates and associated biomarkers on slit lamp images of microbial keratitis.

Methods : The dataset consisted of diffuse, white light, slit lamp images of 120 eyes from University of Michigan Kellogg Eye Center (MI, USA) and Aravind Eye Care System (Madurai, India). The regions of interest (RoIs) were manually segmented by a physician at Michigan using ImageJ software (NIH, MD, USA). Pathological RoIs were stromal infiltrates, white cells (the transition zone of stromal infiltrates), hypopyons, and edema. Non-pathological light reflexes were also segmented as RoIs. A region-based convolutional neural network (R-CNN) was trained with 80 images to detect and segment the RoIs. Data augmentation techniques were used, and training was monitored on a validation set of 20 images. During testing, the trained R-CNN was used to automatically segment images that were not used during training or validation. The Dice similarity coefficient (DSC) between the manual and automatic segmentations was calculated.

Results : Using 6-fold cross-validation, the R-CNN was tested on a subset of 91 eyes for which a second physician’s manual segmentation of stromal infiltrates was available. The DSC (mean ± standard deviation) between manual and automatic segmentations for stromal infiltrates, white cells, hypopyons, edema, and light reflexes was 0.76 ± 0.19, 0.74 ± 0.28, 0.75 ± 0.38, 0.74 ± 0.43, and 0.75 ± 0.24, respectively. Differences often occurred as the boundaries of some RoIs are ambiguous and difficult to segment, even for experienced physicians. As comparison, for manual segmentation of stromal infiltrates, the DSC between the two physicians was 0.81 ± 0.19. Figure 1 shows an example.

Conclusions : Performance of our automatic method is close to human grading although more work needs to be done to further improve accuracy.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.

 

Figure 1: Manual and automatic segmentations of stromal infiltrates (magenta), white cells (cyan), hypopyons (yellow), edema (white), and light reflexes (purple). The manual segmentation of the stromal infiltrate by the second physician is shown in orange.