June 2023
Volume 64, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2023
Post-keratoplasty graft failure detection with Code Free Deep Learning: Image processing and real-life incidence influence on training performance
Author Affiliations & Notes
  • Carlos Mendez Mangana
    Cornea and cataract, Centro de Ojos de La Coruña, A Coruña, Galicia, Spain
    AI engineering, Centre d'Oftalmologia Barraquer, Barcelona, Catalunya, Spain
  • Anton Barraquer Kargacin
    AI engineering, Centre d'Oftalmologia Barraquer, Barcelona, Catalunya, Spain
    Cornea and cataract, Centre d'Oftalmologia Barraquer, Barcelona, Catalunya, Spain
  • Jorge Fernandez Engroba
    AI engineering, Centre d'Oftalmologia Barraquer, Barcelona, Catalunya, Spain
  • Álvaro Ferragut-Alegre
    AI engineering, Centre d'Oftalmologia Barraquer, Barcelona, Catalunya, Spain
  • Pedro Tañá
    AI engineering, Centre d'Oftalmologia Barraquer, Barcelona, Catalunya, Spain
  • Javier Zarranz-Ventura
    Medical Retina and AI, Hospital Clinic de Barcelona, Barcelona, Catalunya, Spain
    Retina, Hospital Clinic de Barcelona Institut Clinic d'Oftalmologia, Barcelona, Catalunya, Spain
  • Gil Santolaria-Rosell
    AI engineering, Centre d'Oftalmologia Barraquer, Barcelona, Catalunya, Spain
  • Maximiliano Olivera
    AI engineering, Centre d'Oftalmologia Barraquer, Barcelona, Catalunya, Spain
    Moorfields Eye Hospital NHS Foundation Trust, London, London, United Kingdom
  • Rafael I Barraquer
    General Director, Centre d'Oftalmologia Barraquer, Barcelona, Catalunya, Spain
    Professor in Ophthalmology, Universitat Internacional de Catalunya, Barcelona, Catalunya, Spain
  • Footnotes
    Commercial Relationships   Carlos Mendez Mangana None; Anton Barraquer Kargacin None; Jorge Fernandez Engroba None; Álvaro Ferragut-Alegre None; Pedro Tañá None; Javier Zarranz-Ventura None; Gil Santolaria-Rosell None; Maximiliano Olivera None; Rafael Barraquer None
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science June 2023, Vol.64, 1103. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Carlos Mendez Mangana, Anton Barraquer Kargacin, Jorge Fernandez Engroba, Álvaro Ferragut-Alegre, Pedro Tañá, Javier Zarranz-Ventura, Gil Santolaria-Rosell, Maximiliano Olivera, Rafael I Barraquer; Post-keratoplasty graft failure detection with Code Free Deep Learning: Image processing and real-life incidence influence on training performance. Invest. Ophthalmol. Vis. Sci. 2023;64(8):1103.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

Purpose : Analyze the influence on training performance of Code-Free Deep Learning 'CF-DL' models for detection of graft failure 'GF' in post-keratoplasty 'pK' patients by image processing (masking, cropping and resolution) on a proprietary dataset.

Methods : A total of 575 images were assembled in a final pK dataset, from a previous study in which we demonstrated the usefulness of CF-DL trained models for GF detection in 220 images (balanced preliminary dataset).
For image processing, OpenCV/Python scripts were developed. An eye contour mask was applied to the complete dataset at native resolution 'NR' (2565x1648) and then the complete dataset was processed in low resolution 'LR' (359x240). The incidence of GF was 21%. Models were trained on Vertex-AI (Google Cloud) AutoML.

Results : The NR model performance without image masking obtained an average precision 'AP' of 0.98. LR models achieved an AP of 0.92, tending to overfitting with 100% detection of ‘healthy’ cases and 66% detection of GF. After masking, NR models showed improved metrics with over 0.98 AP, with 81% detection of GF, whereas LR achieved comparable performance.

Conclusions : In our proof of concept 'POC' study presented at 2022 ARVO, we demonstrated the utility of CF-DL algorithms for GF detection. This POC was based on a balanced dataset. For the current study, classes from the entire dataset were adjusted according to reported incidence of GF in literature.
NR models achieved good performance, comparable to our POC. At LR, an overfitting tendency and poor performance on the clinically relevant form 'GF' was observed. After image masking, NR models kept comparable performance, but LR models drastically improved and achieved metrics comparable to NR. Significant differences in training time were observed, but given the nature of AutoML platform, this remains to be analyzed.
In conclusion, CF-DL trained models with real-life incidence dataset are able to detect graft failure. A data-centric approach, with image manipulation as described, helps to keep relevant model performance even with significant changes in image sizes. Significant impact on training time is evident and even though this cannot be completely analyzed on GCP-AutoML platform, it becomes relevant for fine-tuning our dataset and pipeline of work looking forward to the development of transfer learning and federative learning models.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×