Haochen Zhang

Haochen Zhang

Ph.D. student in ECE at UCSD. Master degree received from USTC in 2020.

Robust AMD Stage Grading with Exclusively OCTA Modality Leveraging 3D Volume

Published in International Conference on Computer Vision (ICCV) Workshop, 2023

In this paper, we experimentally verified that the OCTA projections, which ophthalmologists usually use for diagnosis, are easily affected by layer segmentation errors. Those errors degrade the classification performance.

Interrelationships among OCT and OCTA raw volume, B-scans, and OCTA projection. OCT B-Scan layer segmentation influences OCTA projection generation.
Interrelationships among OCT and OCTA raw volume, B-scans, and OCTA projection. OCT B-Scan layer segmentation influences OCTA projection generation.
Layer segmentation errors, which are prevaling, generate artificts in OCTA projections and do harm to classifier training.
Layer segmentation errors, which are prevaling, generate artificts in OCTA projections and do harm to classifier training.

We propose to use 3D raw OCTA volume to avoid the impacts of those errors. To achieve this, we modify a pretrained 2D network to perform volume classification. We also adopt an additional projection supervision to facilitate training of shallow feature extractor.

The proposed network structures for (a) 2D projections, (b) 3D volumes and (c) 3D volumes with 2D projection supervision. The layers in blue have pretrained weights while those in red are trained from scratch.
The proposed network structures for (a) 2D projections, (b) 3D volumes and (c) 3D volumes with 2D projection supervision. The layers in blue have pretrained weights while those in red are trained from scratch.

Experimental results show that the proposed classifier can achieve the accuracy of more than 80%, regardless of the presence of layer segmentation errors. These results prove the effectiveness of our methods and suggest that OCTA is a promising modality to distinguish various stages of AMD disease.

Experimental results.
Experimental results.

Paper | Vidoe | Poster

Please also refer to our clinical paper

Citation:

@inproceedings{zhang2023robust,
  title={Robust AMD Stage Grading with Exclusively OCTA Modality Leveraging 3D Volume},
  author={Zhang, Haochen and Heinke, Anna and Galang, Carlo Miguel B and Deussen, Daniel N and Wen, Bo and Bartsch, Dirk-Uwe G and Freeman, William R and Nguyen, Truong Q and An, Cheolhong},
  booktitle={IEEE/CVF International Conference on Computer Vision (ICCV) Workshop},
  pages={2411--2420},
  year={2023}
}

@article{heinke2022artificial,
  title={Artificial intelligence for OCTA-based disease activity prediction in age-related macular degeneration.},
  author={Heinke, Anna and Zhang, Haochen and Deussen, Daniel and Galang, Carlo Miguel B and Warter, Alexandra and Kalaw, Fritz Gerald Paguiligan and Bartsch, Dirk-Uwe G and Cheng, Lingyun and An, Cheolhong and Nguyen, Truong and others},
  journal={RETINA},
  pages={10--1097},
  year={2022}
}