Diabetes classification info

[ana-m-m]

Hi! I wanted to know how i can get the dataset with the image paths + biomarkers + patient info + diabetes classification.

In the zenodo download, in the ml_centric_label folder, I cannot find in any of the files diabetes classification, but this info:

Path (Trial/Arm/Folder/Visit/Eye/Image Name),Scan (n/49),Atrophy / thinning of retinal layers,Disruption of EZ,DRIL,IR hemorrhages,IR HRF,Partially attached vitreous face,Fully attached vitreous face,Preretinal tissue/hemorrhage,Vitreous debris,VMT,DRT/ME,Fluid (IRF),Fluid (SRF),Disruption of RPE,PED (serous),SHRM,Eye_ID,BCVA,CST,Patient_ID

But in https://huggingface.co/datasets/gOLIVES/OLIVES_Dataset#data-download , i can get "Disease Label" (float64).

Is there a way to download the complete dataset with Disease Label and image paths, instead of using
olives = load_dataset('gOLIVES/OLIVES_Dataset', 'disease_classification', split = 'train')

Thanks and best regards,

Ana MM

[ana-m-m]

Sorry,just realized that:
PRIME evaluates Diabetic Retinopathy (DR) and TREX-DME evaluates Diabetic Macular Edema
(DME).

[MohitPShankar]

Thanks Ana! Yes, the PRIME and TREX directories store the diseases separately.

We have added a new comment today within the README file on Github to make it clearer.

[ana-m-m]

Thank you for the clarification. I have another question: What does 'OS' and 'OD' stand for in the database?
And, is there any database you know that could be included that includes non-Diabetic, so this classification could be done, using also OLIVES dataset?

[MohitPShankar]

In the appendix of our paper (https://arxiv.org/pdf/2209.11195), we list a number of other datasets in Tables 5 and 6. Please take a look; some of them should be helpful.

OD and OS identify the right or left eye respectively (Page 30 in Appendix, point 5). OD is right eye and OS is left eye

[ana-m-m]

Thanks again for the explanation! I have some questions regarding the training process:

In the Multi-Modal Integration Between OCT and Biomarker/Clinical Labels, for example, when using ResNet-18 as a baseline for classification with only OCT images:

1. Are the OCT images treated as independent samples, each with its own classification (i.e., as individual rows)?
2. Or are they considered as part of the same eye, visit, or patient when training the model?
3. Is the same method applied for all the experiments regarding Multi-Modal Integration Between OCT and Biomarker/Clinical Labels?

Looking forward to your insights!

[MohitPShankar]

They are considered individual samples since the biomarkers do not manifest themselves across all slices in the eye. However, there are papers that have considered each 3D volume as one sample [1]

[1] Oghbaie, Marzieh, et al. "VLFATRollout: Fully transformer-based classifier for retinal OCT volumes." Computerized Medical Imaging and Graphics 118 (2024): 102452.

[ana-m-m]

Hi again! Thank you for all the help.
I want to run some tests of disease classification for OCT, and use yours as baseline for comparison. I would like to use the same train-test split but in Hugging Face data I can just download the train split, not the test. Where could I find it?

(https://huggingface.co/datasets/gOLIVES/OLIVES_Dataset/viewer/disease_classification?views%5B%5D=disease_classification)

Best regards, and thank you again for all the help