Abstract
Purpose :
OCT is rapidly becoming the standard of care for ophthalmic diagnosis. However, ocular abnormalities can be overlooked during clinical examination especially when several images are obtained during a single image capture. Hence, automated detection of ocular abnormalities from OCT images is actively developed for. We propose an abnormality detection method using segmentation of OCT images of normal eyes, and report its effectiveness.
Methods :
We focused on the differences in the difficulty of segmenting OCT images of normal eyes and those containing abnormal regions. An abnormality in OCT images appears as a disrupted retinal layer structure due to disease. When image processing designed for normal retinal layer structure is applied, segmentation of OCT images that contain abnormal portions is considered more difficult than that of OCT images of normal eyes. We propose a model that performs segmentation of OCT images using deep learning. We visualized the degree of abnormality of the retinal structure by quantifying the deviation in the probability distribution output by the proposed model as “uncertainty”. When the model is trained only with OCT images of normal eyes, segmentation of the OCT images containing abnormal portions is difficult, and the uncertainty value is expected increase.
Results :
Twenty OCT images of normal eyes captured with the RS-3000 Advance (Nidek, Aichi, Japan) were used to train the model by deep learning. Then, 32 OCT images of normal eyes and 80 OCT images of diseased eyes were input to the model to obtain the uncertainty values. The obtained uncertainty was 12.70±0.40 for normal eyes, and 16.38±2.09 for diseased eyes. Although the data set was small, the model distinguished between normal eyes and diseased eyes with high accuracy (AUC=0.99). Additionally, there was confirmation that areas with a high uncertainty matched the abnormal regions.
Conclusions :
We present a method to detect retinal abnormalities by quantifying the deviation in the probability distribution output by a model trained by deep learning. This was achieved by focusing on the difference in the difficulty of segmentation that becomes conspicuous between OCT images of normal eyes and those containing abnormal regions. This method can be developed by training the model with OCT images of normal eyes only.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.