Investigative Ophthalmology & Visual Science Cover Image for Volume 64, Issue 8
June 2023
Volume 64, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2023
Detecting sebaceous carcinoma in whole-histopathological slide images using deep learning.
Author Affiliations & Notes
  • Naohiko Funatsu
    Kyushu Daigaku, Fukuoka, Fukuoka, Japan
  • Masato Akiyama
    Kyushu Daigaku, Fukuoka, Fukuoka, Japan
  • Mika Tanabe
    Kyushu Daigaku, Fukuoka, Fukuoka, Japan
  • Hiroshi Yoshikawa
    Kyushu Daigaku, Fukuoka, Fukuoka, Japan
  • Koh-Hei Sonoda
    Kyushu Daigaku, Fukuoka, Fukuoka, Japan
  • Footnotes
    Commercial Relationships   Naohiko Funatsu None; Masato Akiyama NIDEK, Code R (Recipient); Mika Tanabe None; Hiroshi Yoshikawa None; Koh-Hei Sonoda Santen, Code F (Financial Support), HOYA, Code F (Financial Support)
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science June 2023, Vol.64, 3614. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Naohiko Funatsu, Masato Akiyama, Mika Tanabe, Hiroshi Yoshikawa, Koh-Hei Sonoda; Detecting sebaceous carcinoma in whole-histopathological slide images using deep learning.. Invest. Ophthalmol. Vis. Sci. 2023;64(8):3614.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Sebaceous carcinoma (SeC) of the eyelid is a rare and fatal malignant tumor. In case that tumor cells are diffusely scattered in eyelid tissues, it takes a lot of time and effort to determine the area of tumor cells accurately. We aim to distinguish between normal and malignant area in whole-histopathological slide images of SeC using deep learning.

Methods : The 66 whole slide images (WSI) of the eyelids that underwent eyelid tumor resection in the Kyushu University Hospital from July 2012 to August 2019 were retrospectively collected. These WSIs were subdivided into training (N = 50), validation (N = 6), and test data (N = 10). Pathological patches were generated to be 512x512 pixel patches from Hematoxylin-Eosin-stained WSI and were labeled by a trained pathologist and ophthalmologists. Subdivided images were trained by a convolutional neural network. Validation data were used to assess the threshold of predictive value. Area under the receiver operating characteristic curves and F1-score in test data were evaluated to estimate the accuracy of the constructed model.

Results : From all pathology patches (N = 599,113), patches that contained no tissue (background) and less than half tissue were excluded (n = 437,225). Normal (n = 116,433) and malignant (n = 45,455) tissues were divided as follows: normal (n = 92,468) and malignant (n = 27,995) tissues for training data, normal (n = 9,050) and malignant (n = 3,537) tissues for validation data, normal (n = 14,915) and malignant (n = 13,915) tissues for test data. Among the test data, 15,989 (55%) were estimated as normal, and 12,849 (45%) were regarded as the patches including malignant cells by the constructed model. The parameters of accuracy in the results of the test dataset were the overall concordance rate of 92%, an AUC of 0.98, and F1 score of 93%.

Conclusions : Our results confirm that the constructed deep learning model can distinguish malignant tissue from normal in eyelids with SeC.

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.

×