June 2023
Volume 64, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2023
Core-set based batch-mode active learning for intelligent training of optical coherence tomography (OCT) based retinal pathologies detection models
Author Affiliations & Notes
  • Mengze Li
    Carl Zeiss Meditec AG, Munich, Bayern, Germany
  • Abdelrahman Elskhawy
    Carl Zeiss Meditec AG, Munich, Bayern, Germany
  • Krunalkumar Ramanbhai Patel
    Carl Zeiss Meditec AG, Munich, Bayern, Germany
  • Abouzar Eslami
    Carl Zeiss Meditec AG, Munich, Bayern, Germany
  • Ali Salehi
    Carl Zeiss Meditec, Inc., Dublin, California, United States
  • Footnotes
    Commercial Relationships   Mengze Li Carl Zeiss Meditec AG, Code E (Employment); Abdelrahman Elskhawy Carl Zeiss Meditec AG, Code E (Employment); Krunalkumar Ramanbhai Patel Carl Zeiss Meditec AG, Code E (Employment); Abouzar Eslami Carl Zeiss Meditec AG, Code E (Employment); Ali Salehi CarlZeiss Meditec, Inc., Code E (Employment)
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science June 2023, Vol.64, 2393. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Mengze Li, Abdelrahman Elskhawy, Krunalkumar Ramanbhai Patel, Abouzar Eslami, Ali Salehi; Core-set based batch-mode active learning for intelligent training of optical coherence tomography (OCT) based retinal pathologies detection models. Invest. Ophthalmol. Vis. Sci. 2023;64(8):2393.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Advances of deep learning (DL) in medical applications have suffered from lack of sufficient annotated training data due to reasons of confidentiality, time, and cost required to obtain expert-annotated samples. To alleviate this, we use a core-set active learning (AL) approach to effectively query an oracle for the labeling of the most informative and diverse samples.

Methods : To prove the effectiveness of the core-set approach, retrospective OCT study data including retinal B-scans from CIRRUS™ HD-OCT 5000/4000 (ZEISS, Dublin, CA) were analyzed [1, 2]. Each B-scan was labeled by two experts for retinal pathologies with derived binary labels [2]. The training, validation, and testing split is 54784, 7808, and 6400 B-scans corresponding to 428, 61 and 50 cubes respectively. An initial labeled training set of 500 scans was randomly chosen. In each AL iteration, a labeling of size 100 is used. The classification model is SqueezeNet pre-trained on ImageNet. A core-set AL was applied to find a small subset of data that is representative of the distribution. Core-set AL has been proven to be effective in deep active learning when selecting a batch of samples. Performance is compared using classification accuracy against baseline methods namely random sampling, and Bayesian dropout-based uncertainty sampling.

Results : As shown in Fig 1, to achieve the same accuracy, core-set AL queries fewer labeling data compared to random sampling and uncertainty sampling. Uncertainty sampling behaves worse than random sampling due to the batch-mode setting. By only labeling 3,000 B-Scans using core-set AL, we achieved 92% accuracy while a fully supervised model achieved 94% using 54,784 labeled training data.

Conclusions : Core-set AL is an efficient approach to minimize the cost and efforts of labeling medical data. By only selecting the right informative subset to train the DL model on, the process is accelerated, and competitive performance is achieved. Core-set AL also showed potential compared to other methods when combined with CNN models where batch sampling is needed to avoid longer training times. While this holds true for the experiments shown here, more experimentation is needed to show significance of the proposed approach.

[1]Yu et al.IOVS 2020;61(9):PB0085
[2]Ren et al.IOVS 2020;61(7):1635

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

 

Fig 1 Samples to achieve target levels of accuracy

Fig 1 Samples to achieve target levels of accuracy

×
×

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.

×