Unified Cystoscopy Cancer Detection (Dataset F)

Dataset Description

Dataset F is derived from Dataset E by removing every PUNLMP frame and re-deriving a patient-level train / validation / test split from scratch. It is prepared for 3-class deep learning (malignant vs non-malignant vs non-ROI) and keeps the same per-image schema as Dataset E (image, cancer_label, grade_label, subclass_label, source_dataset, original_filename, patient_id, imaging_type, target3, track_id, cv_fold).

• Total images: 14,376 (E had 14,921; 545 PUNLMP frames removed)
• Unique patients / cases: 209 (E had 212; the 3 PUNLMP-only source-B patients 6, 8, 29 are dropped)
• Modalities: White-Light Imaging (WLI), Narrow-Band Imaging (NBI), Blue-Light Cystoscopy (BLC)
• Splits: train / validation / test (70 / 15 / 15), patient-level with no leakage
• 5-fold cross-validation on the train pool (cv_fold 0-4; -1 for val/test)

Relationship to Dataset E

	Dataset E	Dataset F
Total images	14,921	14,376
Patients	212	209
non_malignant images	900	355
PUNLMP frames	545 (source-B non_malignant)	0 (excluded)
Split derivation	seed 16, 14 swaps	seed 17, 12 swaps (re-searched)

PUNLMP Exclusion

In Dataset E, PUNLMP (Papillary Urothelial Neoplasm of Low Malignant Potential) is the histological_type of 12 source-B videos (milkyroad/B), and every source-B non_malignant frame in E comes from those PUNLMP videos (E patients 6, 8, 29; tracks 028, 029, 032-037, 162-165). Therefore:

PUNLMP frames  ==  (source_dataset == 'B') AND (target3 == 'non_malignant')

Removing them drops the non_malignant class from 900 -> 355 images, all now coming from sources C and D. Because this makes a good patient-level split much harder, the split is re-derived from scratch (Dataset F does not inherit Dataset E's split assignment).

Source Datasets

Code	Source	Type	Images in F	Patients in F
B	Cystoscopy video dataset (stride-8 frames)	Video frames	4,555	27
C	CystoDS	Still cystoscopy images	8,067	160
D	Kaggle cystoscopy frames	Still cystoscopy frames	1,754	22

Source-B majority-class frames (malignant + non-ROI) in E are already capped at 60 per (patient, track); Dataset F's split search used cap=None, i.e. it keeps E's existing 60-cap and never adds frames. Source-B non_malignant (PUNLMP) frames are excluded entirely, so all 4,555 source-B images in F are malignant or non_roi.

Patient ID normalization (unchanged from E)

Source	Original ID field	Normalized range	Count
B	patient_id (0-29)	1-30	27 (3 PUNLMP-only patients dropped)
C	pid	31-190	160
D	case_id (0-25)	191-212	22

Label Derivation

Identical to Dataset E. See E's dataset card for the full per-source mapping. The derived 3-class task label is:

subclass_label	target3
malignant (0)	malignant (0)
non_malignant (1)	non_malignant (1)
normal (2)	non_roi (2)
landmark (3)	non_roi (2)
foreign_body (4)	non_roi (2)

Train / Validation / Test Split

Patient-level stratified split using StratifiedGroupKFold (source x dominant-class stratification) for test, then for train/val, followed by greedy patient-swap refinement that minimises the maximum per-class per-split fill-ratio deviation. Hard constraints: every split contains every class and every source; a source-floor of 4 patients per source in val and test.

A grid search over (seed, per-track cap, source_floor) selected the configuration with the lowest score (score = max_fill_dev + penalties):

• Split ratio: 70 / 15 / 15 (train / val / test)
• Best config: cap=None, source_floor=4, seed=17
• Swaps applied: 12
• Max fill-ratio deviation: 0.0047
• No patient appears in more than one split (verified)

Split sizes

Split	Images	Patients
train	10,067	150
validation	2,157	31
test	2,152	28
Total	14,376	209

Class distribution per split

Class	Total	Train	Val	Test
malignant	6,515 (45.3%)	4,562	975	978
non_malignant	355 (2.5%)	249	53	53
non_roi	7,506 (52.2%)	5,256	1,129	1,121

Fill ratios (count / target, ideal = 1.00)

Class	Train	Val	Test
malignant	1.00	1.00	1.00
non_malignant	1.00	1.00	1.00
non_roi	1.00	1.00	1.00

Source x split (patients)

Source	Train	Val	Test
B	19	4	4
C	117	23	20
D	14	4	4

Source x split (images)

Source	Train	Val	Test
B	3,402	447	706
C	5,654	1,224	1,189
D	1,011	486	257

Non-malignant patient coverage

Split	Patients	Images
train	53	249
validation	12	53
test	12	53

All non_malignant images in F come from sources C and D (source-B non_malignant / PUNLMP is excluded).

5-Fold Cross-Validation (train pool only)

StratifiedGroupKFold on the train split (per-image target3 stratification, patient-level grouping, no patient in multiple folds, seed=15, 5 folds).

Fold	Patients	Images	Malignant	Non-mal	Non-ROI
0	27	2,008	912	49	1,047
1	30	2,017	912	50	1,055
2	33	1,988	912	49	1,027
3	29	2,055	913	50	1,092
4	31	1,999	913	51	1,035

The cv_fold column is -1 for validation and test patients.

Note on minority-class (non-malignant) fold balance

After PUNLMP exclusion the non_malignant class shrinks to 355 images and remains heavily patient-concentrated: only 53 train patients carry non_malignant frames and a few patients dominate the count. Because cross-validation is patient-grouped (a patient never appears in more than one fold), high-volume patients cannot be split across folds, so the per-fold non-malignant count varies modestly (49-51) while the majority classes stay near-perfectly balanced (malignant spread 912-913; non-ROI spread 1,027-1,092). Practitioners should report per-fold non-malignant counts alongside metrics and prefer the macro-averaged F1 across all five folds over any single fold's score when estimating minority-class performance.

Dataset Statistics

By target3

Malignant	Non-malignant	Non-ROI
6,515	355	7,506

By subclass label

Malignant	Non-malignant	Normal	Landmark	Foreign body
6,515	355	7,044	211	251

By cancer label

Total	Cancer	Non-cancer
14,376	6,514	7,862

By grade label

High-grade	Low-grade	Not applicable
2,205	4,309	7,862

By imaging modality

WLI	NBI	BLC
13,605	321	450

By source

B	C	D
4,555	8,067	1,754

Features Schema

Column	Type	Description
image	Image	Cystoscopy image (decoded as PIL Image on load)
cancer_label	ClassLabel	non_cancer (0) / cancer (1)
grade_label	ClassLabel	low_grade (0) / high_grade (1) / not_applicable (2)
subclass_label	ClassLabel	malignant (0) / non_malignant (1) / normal (2) / landmark (3) / foreign_body (4)
source_dataset	string	B, C, or D -- original source dataset
original_filename	string	Filename in the original source dataset
patient_id	int32	Normalized sequential patient/case ID (1-212)
imaging_type	ClassLabel	WLI (0) / NBI (1) / BLC (2)
target3	ClassLabel	malignant (0) / non_malignant (1) / non_roi (2) -- 3-class task label
track_id	string	Source-B video track ID (e.g. 008); NA for sources C and D
cv_fold	int64	5-fold CV assignment (0-4) for train patients; -1 for val/test

Loading

from datasets import load_dataset

ds = load_dataset("milkyroad/F")
print(ds)
# DatasetDict({
#     train: 10,067 images
#     validation: 2,157 images
#     test: 2,152 images
# })

sample = ds["train"][0]
print(sample["target3"])  # 0 (malignant), 1 (non_malignant), or 2 (non_roi)

Provenance

Built from milkyroad/E by:

1. Dropping all PUNLMP frames (source_dataset == 'B' and target3 == 'non_malignant').
2. Grid-searching (seed, cap, source_floor) for the patient-level split that minimises the maximum per-class per-split fill-ratio deviation, with hard coverage constraints.
3. Attaching a 5-fold StratifiedGroupKFold CV assignment (seed=15) to the train pool.

Selected config: cap=None, source_floor=4, seed=17, 12 swaps, max fill-ratio deviation 0.0047.