--- pretty_name: RetSAM-Reannotation language: - en task_categories: - image-segmentation size_categories: - n<1K tags: - medical - retinal-imaging - segmentation configs: - config_name: DRIVE data_files: - split: train path: DRIVE/splits/train.jsonl - split: val path: DRIVE/splits/val.jsonl - split: test path: DRIVE/splits/test.jsonl - config_name: HRF data_files: - split: train path: HRF/splits/train.jsonl - split: val path: HRF/splits/val.jsonl - split: test path: HRF/splits/test.jsonl - config_name: GAMMA data_files: - split: train path: GAMMA/splits/train.jsonl - split: val path: GAMMA/splits/val.jsonl - split: test path: GAMMA/splits/test.jsonl - config_name: IDRiD data_files: - split: train path: IDRiD/splits/train.jsonl - split: val path: IDRiD/splits/val.jsonl - split: test path: IDRiD/splits/test.jsonl --- # RetSAM-Reannotation Original fundus images and reannotations associated with [RetSAM](https://arxiv.org/abs/2602.07012), grouped by source dataset. | Dataset | Images | Vessel | OD | OC | EX | HE | |---|---:|---:|---:|---:|---:|---:| | DRIVE | 40 | 40 | 39 | 39 | 16 | 8 | | HRF | 45 | 45 | 45 | 45 | 21 | 38 | | GAMMA | 100 | 100 | 100 | 98 | 16 | 3 | | IDRiD | 81 | 81 | 79 | 79 | 78 | 80 | | **Total** | **266** | **266** | **263** | **261** | **131** | **129** | Annotation columns count images with nonempty labels. OD/OC = optic disc/cup; EX = hard exudates; HE = hemorrhages. MA and SE masks are included but are all empty in this export. **Included:** 266 original images, 266 XML files, 1,354 raw object masks, and 1,596 merged masks (six per image). All image and annotation files retain their original bytes. Each of `DRIVE/`, `HRF/`, `GAMMA/`, and `IDRiD/` contains: - `images/`: original photographs. - `annotations/`: complete source XML and `Masks/` with individual objects, including separate artery and vein labels. - `masks/{vessel,odoc,ex,he,ma,se}/`: merged grayscale masks. - `metadata.jsonl` and `splits/`: image mapping and existing local train/val/test assignments. Paths are relative to the repository root. **Encoding:** OD/OC: 0 = background, 127 = disc excluding cup, 255 = cup; full disc = `mask > 0`. Other merged masks: foreground = `mask > 0`. Raw RGBA object masks: foreground = alpha > 0. **Missing OD/OC:** DRIVE `30_training` and IDRiD `IDRiD_10`, `IDRiD_66` lack both; GAMMA `0024`, `0080` lack OC. Missing annotations are not verified negatives. GAMMA contains 100 public training images; its local splits do not represent hidden challenge test labels. Original image terms remain those of [DRIVE](https://drive.grand-challenge.org/), [HRF](https://www5.cs.fau.de/research/data/fundus-images/), [GAMMA](https://gamma.grand-challenge.org/), and [IDRiD](https://idrid.grand-challenge.org/). No additional annotation license is assigned here.