The CystoFold dataset is the first hand-labeled dataset for tissue fold formation 
prediction in urinary bladder endoscopy. It provides sparse pixel-wise polygon and rectangle annotations capturing 
the transition from smooth to folded bladder tissue across filling and emptying cycles, designed to 
support deformation-aware reconstruction pipelines and SLAM-based surgical navigation systems.



A total of 873 annotated frames were selected from cystoscopic video recordings of 
three patients undergoing routine cystoscopy. Annotations were produced using a custom 
backward-tracking methodology: fold regions are first identified in the emptied 
(folded) bladder state, then tracked back to their appearance in the distended state, with a pinned 
reference view guiding the annotator. All annotations were created using a custom polygon-based GUI tool.



    Important: Sparse Annotation Strategy
    To ensure maximum ground-truth reliability, annotations cover only a small, localized area 
    of the visible tissue. Annotators were instructed to label only regions where the tissue state (Stable, 
    Pre-Fold, or Folded) could be identified with absolute certainty. This "safe-labeling" approach prevents 
    ambiguous tissue from being misclassified.

Semantic Classes

Stable: Tissue that remains flat and visible throughout the emptying cycle
Pre-Fold: Tissue appearing flat in the current frame that will fold before the cycle ends
Folded: Tissue currently forming a visible fold

Cycle Markers

Each video includes explicit markers for the bladder filling cycle. The marker filled 
indicates the frame at which the bladder reached a fully filled state; the marker emptied 
indicates the frame at which it was fully emptied. These markers are provided in the 
_cycles.csv files and allow temporal alignment of annotations with the physiological 
cycle phase. The per-frame filling context (Filling or Emptying) is additionally 
stored in the context field of each annotation record in the JSON and CSV label files.

Dataset Structure

patient_1/ (2 filling cycles, 586 usable frames)

     <video>.mp4 (Original unlabeled video)
        <video>.csv (Per-polygon annotation records)
        <video>_cycles.csv (Cycle phase marker timestamps)
        <video>.json (Full labeler session, restorable)


patient_2/ (1 filling cycle, 103 usable frames)

        ...


patient_3/ (1 filling cycle, 184 usable frames)

        ...

Usage

Each annotation record in the .csv file corresponds to a single labeled region in one frame.
The relevant columns are:


frame — video frame index (0-based)
label — class name: Stable (1), Pre-Fold (2), Folded (3); Background is implicitly 0
shape — annotation geometry: polygon (precise boundary, polygon_points field contains a JSON array of [x, y] pairs) or rect (bounding box, use x1, y1, x2, y2 columns)
use_frame — set to False for frames that should be excluded; filter to use_frame != False before training
frame_quality — exclude frames marked motion_blur
context — Filling or Emptying, indicating the bladder phase at that frame


To render a segmentation mask for a given frame, iterate over all rows matching that frame index, 
draw each polygon or rectangle onto a blank canvas using the class index as fill value, and apply 
the result as a per-pixel label map.