Investigative Ophthalmology & Visual Science Cover Image for Volume 61, Issue 9
July 2020
Volume 61, Issue 9
Free
ARVO Imaging in the Eye Conference Abstract  |   July 2020
Multi-label multi-task macular Optical Coherence Tomography (OCT) scans classification system
Author Affiliations & Notes
  • Krunalkumar Ramanbhai Patel
    CARIn, Carl Zeiss India, Bangalore, Karnataka, India
  • SANDIPAN CHAKROBORTY
    CARIn, Carl Zeiss India, Bangalore, Karnataka, India
  • Ganesh Babu
    CARIn, Carl Zeiss India, Bangalore, Karnataka, India
  • Footnotes
    Commercial Relationships   Krunalkumar Ramanbhai Patel, Carl Zeiss India (Bangalore) Pvt. Ltd. (E); SANDIPAN CHAKROBORTY, Carl Zeiss India (Bangalore) Pvt. Ltd. (E); Ganesh Babu, Carl Zeiss India (Bangalore) Pvt. Ltd. (E)
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science July 2020, Vol.61, PB0097. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Krunalkumar Ramanbhai Patel, SANDIPAN CHAKROBORTY, Ganesh Babu; Multi-label multi-task macular Optical Coherence Tomography (OCT) scans classification system. Invest. Ophthalmol. Vis. Sci. 2020;61(9):PB0097.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Macular OCT cubes generally contain multiple B-scans, making the clinical decision time-consuming and tedious. For an effective and fast clinical decision, it is essential to develop an automatic algorithm that identifies the anomalous cube and B-scans, the pathologies present within the B-scans, and overall disease. We present an algorithm that predicts whether the B-scan contains abnormality or not along with the presence of multiple pathologies and also the cube level abnormalities.

Methods : The multi-label multi-task algorithm is developed in two phases using 1485 OCT macular cubes (75,264 B-scans) of 328 subjects (132 normal, 196 with macular pathologies) acquired using 512x128 and 512x32 cube scans from CIRRUS™ HD-OCT 5000 (ZEISS, Dublin, CA) and PRIMUS 200 (ZEISS, Dublin, CA), respectively. In phase one, the trained multi-label Inception-V1 DNN classifies whether the input B-scan is abnormal or not and also predicts the multiple pathologies present in the B-scan. In phase two, for the cube level classification, Principal Component Analysis was performed on the multi-pathologies prediction scores for all B-scans of the cube for dimensionality reduction. The reduced features were then used as inputs for an SVM classifier to classify whether the cube contains abnormality or not. Figure 1 shows the block diagram of the proposed system. The performance of the algorithm is evaluated in terms of Area Under the Curve (AUC) using an independent test set (339 OCT cubes of 82 subjects consist of 17,664 B-scans) and described in result section.

Results : As shown in figure 2 (a), we achieved AUC of 93.71% for the B-scan level binary classification; as shown in figure 2 (b), the algorithm achieved 94.50% AUC for the cube level binary classification and as shown in figure 2 (c), the multi-label classification achieved highest AUC of 98.28% for Vitreo Macular Traction and lowest AUC of 77.95% for Hyper Reflective Particles.

Conclusions : We proposed a multi-label multi-task classification algorithm that classifies not only the B-scan abnormality but also the pathologies present in the B-scan as well as the cube level abnormality classification. As a next step, the work will be extended for Cube level multiple disease classification.

This is a 2020 Imaging in the Eye Conference abstract.

 

Figure 1. Block diagram of proposed multi-label multi-task OCT classification algorithm

Figure 1. Block diagram of proposed multi-label multi-task OCT classification algorithm

 

Figure 2. AUC curves: Multi-label multi-task algorithm evaluation

Figure 2. AUC curves: Multi-label multi-task algorithm evaluation

×
×

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.

×