The participants included people with DME or “1” (73 subjects), Healthy or “0” (54 subjects), and Non-diabetic patients or “2” (64 subjects) with AMD, CNV, or Macular Hole (MH) disorders. Each subject contains 300 B-scans with 300 to 1200 x 300 pixels resolution. These 191 are available in “1-Raw Dataset”.
Since some of the data have poor quality or the macular region was not captured, Quality Assessor (QA) was designed to automatically determine if the B-scan is suitable for further analyses or not. For the data preparation step, the manual quality assessment was performed on 191 volumes. The data was divided into two categories: 190 folders of qualified B-scans and 188 folders of non-qualified B-scans. Each folder consists of 1 to 300 B-scans. And all B-scans in each folder belong to one specific subject. The training data for QA is available in “0-QA-data”. Deep Convolutional Neural Network 1 method is employed to train QA. In the next step, QA was used to assess the quality of subjects in “1-Raw Dataset”. Volumes with more than 30 qualified B-scans passed.
0-QA-data: This data includes 190 folders of qualified B-scans and 188 folders of non-qualified B-scans. Each folder consists of 1 to 300 B-scans with 300 x 300 pixels resolution.
1-Raw Dataset: This dataset includes all 191 volumes before any processing. Each volume contains 300 B-scans with 300 to 1200 x 300 pixels resolution. The participants included people with DME or “1” (73 subjects), Healthy or “0” (54 subjects), and Non-diabetic patients or “2” (64 subjects) with AMD, CNV, or Macular Hole (MH) disorders.
2-Raw Dataset QA: Includes 161 volumes which has been obtained after image quality assessment with automatic cropping. Each subject contains 30 to 300 B-scan with 300 x 300 pixels resolution. The participants included people with DME or “1” (64 subjects), Healthy or “0” (50 subjects), and Non-diabetic patients or “2” (47 subjects) with AMD, CNV, or Macular Hole (MH) disorders.
3-Enhanced Dataset QA: Includes 161 volumes which has been obtained after image contrast enhancement using Gaussianization transformation, and normalization on “2-Raw Dataset QA”. The participants included people with DME or “1” (64 subjects), Healthy or “0” (50 subjects), and Non-diabetic patients or “2” (47 subjects) with AMD, CNV, or Macular Hole (MH) disorders.
4-Denoised dataset QA: Includes 161 volumes which has been obtained after image denoising2, and normalization on “3-Enhanced Dataset QA”. The participants included people with DME or “1” (64 subjects), Healthy or “0” (50 subjects), and Non-diabetic patients or “2” (47 subjects) with AMD, CNV, or Macular Hole (MH) disorders.
5-Aligned Dataset QA: Includes 161 volumes which has been achieved after image alignment based on Internal Limiting Membrane (ILM) using the semi-automatic method, and normalization on “4-Denoised Dataset QA”. The participants included people with DME or “1” (64 subjects), Healthy or “0” (50 subjects), and Non-diabetic patients or “2” (47 subjects) with AMD, CNV, or Macular Hole (MH) disorders.
[1] J. Kim, A. D. Nguyen, and S. Lee, “Deep CNN-Based Blind Image Quality Predictor,” IEEE Trans. Neural Networks Learn. Syst., vol. 30, no. 1, pp. 11–24, 2018.
[2] Z. Amini, and H. Rabbani, “Optical coherence tomography image denoising using Gaussianization transform,” J. Biomed. Opt., vol. 22, no. 8, 2017.
Download Data:
Please cite the following papers to use the data:
- R. Rasti, R. Rostamian, A. Hamidi, A. Zam, M. Ommani, A. Dehghani, and H. Rabbani, “B-Net: A New Strategy for Quality Control and Classification of Low Contrast OCT Images” (submitted to Biomedical Optics Express)