| --- |
| license: cc-by-4.0 |
| task_categories: |
| - image-segmentation |
| tags: |
| - medical |
| - retinal |
| - fundus |
| - vessel-segmentation |
| - ophthalmology |
| size_categories: |
| - n<1K |
| --- |
| |
| # CHASE_DB1 |
| |
| Retinal blood vessel segmentation from color fundus photographs, collected in the |
| **Child Heart and Health Study in England (CHASE)**. |
| |
| ## Contents |
| |
| - **28 images** = 14 children x 2 eyes (left/right) |
| - **999 x 960** RGB fundus photographs (Nidek NM-200-D handheld camera, 30 deg FOV) |
| - **Two independent manual vessel annotations per image** |
| |
| | Column | Description | |
| |---|---| |
| | `image` | RGB fundus photograph | |
| | `mask` | **1st human observer (1stHO) - the gold-standard ground truth.** Binary, `0` = background, `1` = vessel | |
| | `mask_2ndHO` | 2nd human observer. Binary, same encoding. **A human-performance baseline, not the GT** | |
| | `image_id` | e.g. `Image_01L` | |
| | `subject_id` | e.g. `01` - the grouping key (see split caveat below) | |
| | `eye` | `L` or `R` | |
|
|
| ### Which mask is the ground truth? |
|
|
| Use **`mask` (1stHO)**. This is the tier that Fraz et al. (2012) and effectively all |
| subsequent CHASE_DB1 literature train and evaluate against. The second observer's |
| annotations exist so papers can report a "human observer" row. |
| |
| Measured on this archive: mean vessel fraction is **6.93% (1stHO)** vs **6.64% (2ndHO)**, |
| and **inter-observer Dice is 0.777** (range 0.739-0.827). That ~0.78 is the effective |
| human ceiling on this dataset - a model scoring below it is not necessarily failing. |
| |
| ## Splits |
| |
| The official archive is **completely flat and ships no split.** This upload applies the |
| **community-standard 20 train / 8 test** convention (the "first 20 images" split), which |
| in sorted order works out to subjects 01-10 for training and subjects 11-14 for testing. |
| |
| **Split caveat:** each child contributes two eyes, so a random 28-way split leaks fellow |
| eyes across train and test. Always group by `subject_id`. The split above is |
| **subject-disjoint** by construction. |
|
|
| Two other conventions appear in the wild and are *not* used here - be careful when |
| comparing numbers across papers: |
| - Idiap `bob`/`mednet` uses **8 train / 20 test** (reversed) |
| - At least one HuggingFace mirror uses a 14/14 split |
|
|
| ## Field-of-view masks |
|
|
| **None.** CHASE_DB1 ships no official FOV/circular masks. FOV masks distributed with some |
| third-party packages are generated by those packages, not by the dataset authors, so no |
| `fov_mask` column is included here rather than passing synthetic annotation off as |
| dataset content. |
|
|
| ## Provenance |
|
|
| The original Kingston University download pages are no longer reachable |
| (`blogs.kingston.ac.uk/retinal/chasedb1/` no longer resolves; `researchdata.kingston.ac.uk` |
| has a broken DNS delegation; the `staffnet` mirror now returns 403). This upload was |
| built from the archived copy of the official EPrints record, and its integrity was |
| confirmed by checking it byte-for-byte against the independently archived copy from the |
| second official Kingston host - **both are identical** (`md5 d9e47c4bac125b29996fae4380a68db1`, |
| `sha256 bb750b73633492d700e56d91578ee3cdb7e4dbd08279e7465583e9efd26790b6`). The 28-image |
| count matches the count reported in the paper. |
|
|
| The only transformation applied is re-encoding the 1-bit mask PNGs as 8-bit PNGs with |
| values `{0, 1}`; image pixels and mask geometry are unchanged. |
|
|
| ### Note on a common misstatement |
|
|
| Several secondary sources describe CHASE_DB1 as *"56 images from 28 subjects (14 healthy, |
| 14 diabetic retinopathy)."* This is **incorrect.** It is 28 images from 14 healthy |
| schoolchildren; there is no diabetic-retinopathy cohort. |
| |
| ### Resolution |
| |
| The authors quote 1280x960 as the camera's capture resolution, and many papers repeat it, |
| but every distributed file - all 28 images and all 56 masks - is **999x960**. |
| |
| ## Overlap with other retinal datasets |
| |
| None known. CHASE_DB1 is a UK schoolchildren cardiovascular cohort, disjoint in population, |
| country, camera, and era from DRIVE (Netherlands), STARE (US), HRF (Germany), IDRiD (India), |
| PAPILA (Spain), and RAVIR (infrared SLO). RITE is derived from DRIVE, not CHASE. |
|
|
| ## License |
|
|
| **CC BY 4.0.** Copyright Kingston University and St George's, University of London. |
| Redistribution is permitted with attribution. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @article{fraz2012ensemble, |
| title = {An Ensemble Classification-Based Approach Applied to Retinal Blood Vessel Segmentation}, |
| author = {Fraz, Muhammad Moazam and Remagnino, Paolo and Hoppe, Andreas and |
| Uyyanonvara, Bunyarit and Rudnicka, Alicja R. and Owen, Christopher G. and |
| Barman, Sarah A.}, |
| journal = {IEEE Transactions on Biomedical Engineering}, |
| volume = {59}, |
| number = {9}, |
| pages = {2538--2548}, |
| year = {2012}, |
| doi = {10.1109/TBME.2012.2205687} |
| } |
| ``` |
|
|
