Abstract
Purpose :
Determination of IR image quality is paramount to effectiveness of eye motion tracking algorithms which enables reduced motion in optical coherence tomography (OCT) image acquisition. In this project we create a data driven approach for IR image quality classification with minimal manual labeling effort for real time use.
Methods :
We collected 9659 IR images from 8 subjects for 9 different fixations including central and peripheral fixations and generated a dataset of acceptable quality (AQ) and unacceptable quality (UQ) IR images. We use our reference image-based IR tracking algorithm to generate the training data.
The tracking algorithm, which relies on generating landmark points in two images and comparing them, was used to create rules for classifying AQ & UQ images. The first IR tracking image is used as a reference image for each fixation and it is reviewed manually if necessary. The tracking output landmark number and distribution were used as a measure of IR moving image quality to generate initial training set divided into the AQ & UQ classes.
A grader quickly reviews the initial training set to correct misclassifications in the training set (Fig1). The input to the neural network are 3 adjacent temporal frames of the IR images, and a sequence is labelled AQ only if all three frames are of AQ quality according to the pipeline output. Finally, a VGG style network is trained on this dataset to predict the quality of given image. After an initial round of training the network is run on a large UQ image dataset created using the pipeline. The 100 highest confidence incorrect images are included in the train set and the network is retrained using these hard negative images.
The final network is run on an independent holdout test set of 3 subjects with various fixations.
Results :
The resulting performance is measured by visualizing the images as a video sequence along with the continuous confidence prediction curve (Fig 2). The network has an area under the curve (AUC) of 0.905 on the hold out test set along with 40 ms runtime on an Intel CoreTM i7-9870H CPU.
Conclusions :
The initial results of the proposed training method indicate that real time image quality assessment for IR image has reasonable performance with the limited number of available datasets which enables usage in a real-world setting with minimal labelling effort.
This is a 2021 Imaging in the Eye Conference abstract.