Fundus Photograph Discrimination Using Transfer Learning over Limited Computing Power Environment
2
조회
0
다운로드
1
인용
0
참고문헌
키워드
초록
In this paper, we demonstrates a reliable and efficient approach to detect eye-related disease with automated fundus screening using convolutional neural network (CNN) and transfer learning that counteracts to insufficient annotated data set and image domain shifts. The weight values learned from the data sets can be used as initial parameters for the other desired neural networks, and additional learning can be conducted on top of the pre-learned model, called transfer learning. It is a particularly useful method when the number of data sets is and small over limited computing power environment. Four different fundus image data sets, image domains such as ethnicity of the target and equipment which the fundus photograph was captured with, were used for the validation. The data sets were annotated by ophthalmologists as healthy, abnormal, or diabetic retinopathy. The ResNet-18 model, pre-trained with ImageNet data set of 1.2 million images of 1000 daily routine objects, were used for transfer learning. The pre-trained model were modified and additionally trained to learn features from the fundus images, and were validated with separate test sets. Given limited quantity of fundus photograph data set and various image domains, the deep learning models can yield robust ophthalmological performance in discriminating pathologies in the eyes. In spite of the simplicity, this study illustrates the capability of transfer learning and suggests pragmatic and practical approach to varied medical settings with fluctuating status of data maintenance and different image domains.
Abstract
In this paper, we demonstrates a reliable and efficient approach to detect eye-related disease with automated fundus screening using convolutional neural network (CNN) and transfer learning that counteracts to insufficient annotated data set and image domain shifts. The weight values learned from the data sets can be used as initial parameters for the other desired neural networks, and additional learning can be conducted on top of the pre-learned model, called transfer learning. It is a particularly useful method when the number of data sets is and small over limited computing power environment. Four different fundus image data sets, image domains such as ethnicity of the target and equipment which the fundus photograph was captured with, were used for the validation. The data sets were annotated by ophthalmologists as healthy, abnormal, or diabetic retinopathy. The ResNet-18 model, pre-trained with ImageNet data set of 1.2 million images of 1000 daily routine objects, were used for transfer learning. The pre-trained model were modified and additionally trained to learn features from the fundus images, and were validated with separate test sets. Given limited quantity of fundus photograph data set and various image domains, the deep learning models can yield robust ophthalmological performance in discriminating pathologies in the eyes. In spite of the simplicity, this study illustrates the capability of transfer learning and suggests pragmatic and practical approach to varied medical settings with fluctuating status of data maintenance and different image domains.
참고문헌 (25)
- V. Biousse, "Ophthalmoscopy in the 21st century: The 2017 H. Houston merritt lecture", Neurology, 90(4), pp.167~175 (2017)
- N. Irani, "Feasibility of a nonmydriatic ocular fundus camera in an outpatient neurology clinic", The Neurologist, 25(2), pp.19~23 (2020)
- Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care office, "M. D. Abràmoff", ~, pp.NPJ Digital Medicine 1, No. 39 (2018)
- A. Grzybowski, "Artificial intelligence for diabetic retinopathy screening: A review", Nature: Eye, 34(3), pp.451~460 (2020)
- P. Heydon, "Prospective evaluation of an artificial intelligence-enabled algorithm for automated diabetic retinopathy screening of 30,000patients", British Journal of Ophthalmology, 105(5), pp.723~728 (2021)
- National institute of food and drug safety evaluation: newly developed medical device and outlook analysis report, "Ministry of Food and Drug Safety of Korea", 12~45 (2020)
- P. Keane, "Reinventing the eye exam", The Lancet, 394(10215), pp.2141~ (2019)
- J. D. Fauw, "Clinically applicable deep learning for diagnosis and referral in retinal disease", Nature Medicine, 24(9), pp.1342~1350 (2018)
- S. J. Pan, "A survey on transfer learning", IEEE Transactions on Knowledge and Data Engineering, 22(10), pp.1345~1359 (2009)
- A. Kashani, "Past, present and future role of retinal imaging in neurodegenerative disease", Progress in Retinal and Eye Research (2021)
- E. Golombievski, "Frequency of direct funduscopy upon initial encounters for patients with headaches, altered mental status, and visual changes: a pilot study", Frontiers in Neurology, 6(233), pp.233~ (2015)
- D. D. Mackay, "The demise of direct ophthalmoscopy: a modern clinical challenge", Neurology: Clinical Practice, 5(2), pp.150~157 (2015)
- Y. LeCun, "Deep learning", Nature, 521(7553), pp.436~444 (2015)
- A. G. Schwing, "Fully connected deep structured networks", ~ (2015)
- M. D. Zeiler, "Visualizing and understanding convolutional networks", 2014European Conference on Computer Vision, pp.818~833 (2014)
- 김백기, "시계열 데이터를 기반으로 한 CNN의 성능 향상 방법에 대한 비교 연구", 한국지식정보기술학회 논문지, 16(5), pp.935~944 (2021), 한국지식정보기술학회[DOI]
- 신승연, "의생명 이미지 분류 기법의 최신 동향", 한국지식정보기술학회 논문지, 12(5), pp.765~776 (2017), 한국지식정보기술학회[DOI]
- K. He, "Deep residual learning for image recognition", 2016IEEE Conference on Computer Vision and Pattern Recognition, pp.770~778 (2016)
- T. A. Liu, "Deep learning and transfer learning for optic disc laterality detection: implications for machine learning in neuro-ophthalmology", Journal of Neuro-Ophthalmology, 40(2), pp.178~184 (2020)
- S. Sengupta, "Ophthalmic diagnosis using deep learning with fundus images – A critical review", Artificial Intelligence in Medicine, 102, pp.101758~ (2020)
- A. Budai, "Robust vessel segmentation in fundus images", International Journal of Biomedical Imaging, 2013 (2013)
- A. D. Hoover, "M. Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response", IEEE Transactions on Medical Imaging, 19(3), pp.203~210 (2000)
- G. Quellec, "Automatic detection of rare pathologies in fundus photographs using few-shot learning", Medical Image Analysis, 61 (2020)