---
license: cc-by-sa-4.0
task_categories:
- image-classification
tags:
- medical
- cystoscopy
- bladder-cancer
size_categories:
- 1K<n<10K
---

# Dataset G — Cystoscopy Lesion Benchmark (Domain-Shift Protocol)

Frame-level cystoscopy dataset for the 3-class task **MT / NML / NROI**
(malignant tumor / non-malignant lesion / non-ROI), built for the
domain-shift benchmark protocol: **train = Sources B+C, test = Source D (TURBT)**.

## Splits (patient-disjoint, frozen)
- `train`: 7,792 frames / 127 patients (Sources B+C)
- `val`: 1,504 frames / 28 patients (hardcoded VAL_PATIENTS, greedy-search selected)
- `test`: 1,754 frames / 22 patients (**Source D only** — frozen forever for the domain-shift track)

Test class support: MT 1,116 / NML 134 / NROI 504. Test contains 1,433 WLI + 321 NBI frames; train has no NBI.

## Columns
`image` (raw letterbox frames), `cancer_label`, `grade_label`,
`subclass_label`, `source_dataset`, `original_filename`, `patient_id`,
`imaging_type` (0=WLI, 1=NBI, 2=BLC — inferred), `target3`
(0=MT, 1=NML, 2=NROI), `track_id`, `lesion_id`, `split`, `cv_fold`
(frozen 5-fold StratifiedGroupKFold, seed 15; val frames carry -1).

## License
Mixture of CC BY 4.0 (C: CystoDS; D: Lazo) and CC BY-SA 4.0 (B) sources;
the dataset ships as **CC BY-SA 4.0**.

## Intended use
Pre-registered domain-shift benchmark: the test split must remain frozen;
do not train on it. Lesion-level evaluation via `lesion_id` is recommended.