News

Video

ARVO 2024: Deep learning model for GA segmentation, adaptable to SS-OCT and SD-OCT data

At this year's ARVO meeting, Qinqin Zhang, PhD, presented a poster titled "A unified deep learning model for geographic atrophy segmentation: Adaptable to SS-OCT and SD-OCT data with multiple scan patterns." At the conference she gave Ophthalmology Times an overview.

At this year's ARVO meeting, Qinqin Zhang, PhD, presented a poster titled "A unified deep learning model for geographic atrophy segmentation: Adaptable to SS-OCT and SD-OCT data with multiple scan patterns." At the conference she gave Ophthalmology Times an overview.

Video Transcript:

Editor's note: The below transcript has been lightly edited for clarity.

Qinqin Zhang, PhD:

Hi, my name is Qinqin Zhang. I'm a senior application data scientist at Carl Zeiss Meditec. So thank you for giving me this opportunity to present our work. So at ARVO in Seattle, I'm presenting a paper titled, "A unified deep learning model for geographic atrophy segmentation: Adaptable to SS-OCT and SD-OCT data with multiple scan patterns" So, in this paper, we will describe a deep learning model combined with the OCT technique to accomplish automatic and the precise assessment or deletion. So, the goal was to train a deep learning model that can adjust the micro scan patterns or spectral domain OCT and the swept-source OCT images to automatically and reliably segment GA. So, to achieve this goal, we extracted 3 features from the oct volume data. So, what is [the] retinal thickness value that can detect the changes from retinal layers and the other is the optical attenuation coefficient values that can detect the abnormality IP layers and the sub IP layers that can capture the hyper transmission defects to detect the other attendees they incur the layers. So by combining these 3 features, a color image was generated and fed into the AI model. So we tune the hyper meter audit model based on the ground truth generated by the clinical expert. So then we can get automatically segmentation on the GA and then mirror size. The results of this research achieved a high sensitivity of 0.99, and a high specificity of 0.93 on eye level detection or the GA with a high DSC of 0.92 compared to the ground truths. That is indicating a high accuracy in segmenting GA. So this is exciting, because this model the holds promise as a valuable management tool for GA segmentation, monitoring, and also detection. So with adaptability with respect to spectral domain OCT and swept-source, OCT and other scan patterns. So, we are excited to present at ARVO and hopefully that will be beneficial to the attendance here. Thank you.

Related Videos
© 2024 MJH Life Sciences

All rights reserved.