Purchase this article with an account.
Bhavna J Antony, Stefan Maetschke, Rahil Garnavi; Deep-learning Framework for Summarization of OCT Volumes. Invest. Ophthalmol. Vis. Sci. 2018;59(9):1226.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
: To develop, validate and test a deep-learning based approach for the summarization of OCT volumes via classification of B-scans as normal or abnormal (with discernible structural defects).
The method was trained and tested using 115 OCT volumes from one eye of elderly patients control subjects and 130 volumes acquired from eyes with intermediate AMD (AREDS2 study) scanned on a Bioptigen OCT scanner (Leica Microsystems Inc.). The macula-centred volumes were acquired from a 6.7mmx6.7mmx2mm region and each B-scan contained 512x1000 pixels. Each of the 100 B-scans per AMD volume were manually labelled as normal or abnormal based on whether visibly discernible defects existed. The training and validation sets consisted of 17,852 and 4476 frames, respectively, with 3925 and 1126 abnormal frames in the two sets, respectively. The remaining B-scans (2154) were not utilized as they were found to be of insufficient quality. A transfer learning approach was utilized beginning with the 16-layer VGGNet – a commonly used deep-learning framework. The last two layers of the model were replaced with two dense layers with 1024 nodes each (rectilinear activation), and a final layer of two nodes (softmax activation), and were the only ones retrained. The images were downsampled to 256x256 before being processed. A balanced cross entropy loss with the Adam optimizer was used during training. For purposes of analysis, the abnormal scans in the testing set were also further divided into mild or severe where small isolated drusen (RPE-drusen complex > 100 mm) were classified as mild, while geographic atrophy, fluid-filled regions and the presence large drusen (RPEDC > 140 mm) were classified as severe (see Fig. 1). The overall accuracy, sensitivity and specificity of the detection of abnormal scans were computed on the testing set.
The overall accuracy, sensitivity and specificity on the validation set was 0.93, 0.83 and 0.97, respectively. Mild abnormalities accounted for 75% of the missed abnormal slices. The classification of an entire volume using this approach takes 5.69 seconds.
The proposed system efficiently classifies B-scans from OCT volumes. This tool shows promise for the summarization of OCT volumes, where specific abnormal B-scans and their location can be quickly identified to assist in diagnosis and patient management.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.
Fig.1. Example of mild (top) and severe (bottom) B-scans.
This PDF is available to Subscribers Only