Abstract
Purpose :
To develop a semantic segmentation algorithm highlighting anatomical structures of interest in digital gonioscopic images acquired by a semi-automatic ophthalmic device, and to overcome ground truth limitations (e.g. missing or incomplete target annotations) by adaptive identification of the most informative region of interest (ROI) in the data and estimation of spatial prediction uncertainty.
Methods :
In gonioscopic acquisitions, only the central part of each image is well illuminated, and the focus varies according to the selected focal plane; therefore, only part of the image can be reliably annotated. A dataset of 274 irido-corneal sector images have been annotated by four experienced ophthalmologists and used to train (202), validate (41) and test (31) a custom Dense-Unet from scratch. The network is trained to achieve two simultaneous, complementary aims: exploiting the ground truths to maximise the segmentation accuracy within the annotated region of the images, and learning to evaluate the informative value of local regions based mainly on sharpness and lightness, locating a ROI. The ROI is then used to filter out uncertain segmentation outputs. Moreover, the use of drop-out during inference makes it possible to generate multiple predictions, enabling us to estimate uncertainty via the pixel-wise variance of the predicted probabilities.
Results :
With a test set representative for relevant clinical cases and un-correlated with the training and validation ones we obtain an overall pixel-wise classification accuracy above 90% within the annotated area of the ground truth data. The automatic ROI identification can locate the most informative region in every test image and the uncertainty estimation proves to effectively highlight most of the un-correctly predicted image pixel sub-sets.
Conclusions :
The proposed system can maximise the information learnt from ground truth annotations and combine it with an effective ROI localization to provide an accurate segmentation of irido-corneal angle layers. Uncertainty estimation can help for a better interpretation of model predictions and may act as an important support in clinical applications.
This is a 2021 ARVO Annual Meeting abstract.