Dataset

[G-qingqian]

Can your dataset(RSNA-BSD1K) ensure that the data are independently and identically distributed by patient? That is, there will be no data from the same patient in both the training set and the validation set.

[Tajamul21]

Yes, the RSNA-BSD1K dataset is IID at the patient level. We ensure that all samples belonging to the same patient are assigned exclusively to a single split, so there is no overlap between training, validation, or test sets. This prevents information leakage and guarantees proper independent evaluation.

[G-qingqian]

Thank you for your reply! However, I noticed that images like 5878_275934305.png and 5911_1491483248.png appear repeatedly in both the full folder and the val folder. The image 32801_1368532850.png in the full folder is very similar to 32801_1368532850.png in the val folder, and it seems like they come from the same patient's data. ---- Replied Message ---- FromTajamul ***@***.***>Date12/16/2025 ***@***.***>***@***.***>, ***@***.***>SubjectRe: [Tajamul21/D-MASTER] Dataset (Issue #7) Tajamul21 left a comment (Tajamul21/D-MASTER#7) Yes, the RSNA-BSD1K dataset is IID at the patient level. We ensure that all samples belonging to the same patient are assigned exclusively to a single split, so there is no overlap between training, validation, or test sets. This prevents information leakage and guarantees proper independent evaluation. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[Tajamul21]

Dear @G-qingqian,

These images are different views (CC vs. MLO), so they naturally look similar but are not duplicates. Cancer can appear differently across views, and each view is treated as a separate image. The splits remain IID because all views from the same patient stay within the same split during actual training/validation.