tbaumgauntlet/parallax-brain-loc-data

Parallax brain-tumor localization data

Single contrast-enhanced T1 brain-MRI slices with one tumor bounding box each on a 0–1000 grid ({"bbox":[x1,y1,x2,y2]}, origin top-left, x right, y down), derived from the manual tumor masks of Jun Cheng's Brain tumor dataset (figshare, doi:10.6084/m9.figshare.1512427.v8, CC BY 4.0). Research use only.

Attribution and modifications

Source images and masks: Jun Cheng, Brain tumor dataset, figshare (2017), doi:10.6084/m9.figshare.1512427.v8, CC BY 4.0. Please also cite Cheng et al., PLoS ONE 10(10): e0140381 (2015) and PLoS ONE 11(6): e0157112 (2016).

Modifications (this repository): slices decoded from the original .mat files in MATLAB-native orientation (never the transposed Hugging Face mirror); rendered as 512×512 8-bit grayscale PNGs with a per-slice linear window [min, p99.9] (no crop or resize); each mask converted to its enclosing box with outward rounding; patient-family-disjoint splits; selection and annotation-quality review; teacher-generated auxiliary text (below). Nothing else is newly authored.

Splits (patient-family disjoint; Phase 08 partition manifests/partitions.json)

File	Images	Use
train_curriculum.jsonl	1024	nested training curriculum; row k belongs to every N ≥ k (positions 1–128 = data v2)
official_holdout.jsonl	54	final evaluation: one slice per official family (18 per class)
development_pilot.jsonl	30	development set (used to diagnose failures and design v2)
fewshot.jsonl	3	the frontier few-shot demonstrations

Each row: image path and sha256, class label (metadata only; never a model input), patient family and PID, source .mat slice id and sha256, mask sha256, pixel box, normalized box, size stratum, review fields.

Training files (training/, MLX-VLM format)

Run	Primary rows	Teacher aux rows	Families	Classes	Unreviewed slices
N128	128	42	78	{'MENINGIOMA': 32, 'GLIOMA': 32, 'PITUITARY': 64}	0
N256	256	42	99	{'MENINGIOMA': 75, 'GLIOMA': 75, 'PITUITARY': 106}	86
N512	512	42	99	{'MENINGIOMA': 160, 'GLIOMA': 160, 'PITUITARY': 192}	311
N1024	1024	42	99	{'MENINGIOMA': 340, 'GLIOMA': 339, 'PITUITARY': 345}	755
v2_noteacher	128	0	—	—	—

At N = 512 and 1,024 the same 99 training families supply many slices each: 1,024 slices are not 1,024 patients.

Teacher-generated data

teacher/accepted.jsonl (42 rows): short descriptions of visible location cues for pituitary training slices, written by anthropic/claude-sonnet-5.5 from the unannotated slice plus the human reference box, filtered by deterministic rules and a blind openai/gpt-6.1-sol review. All raw generations, reviewer replies and accept/reject reasons are in teacher/. The teacher text is model-generated and not radiologist-verified; the boxes are human annotations.

Annotation quality

Every v2 slice and every official slice had a blind review (reviews/); larger curriculum sizes use staged review (automatic anomaly flags, review of flagged slices, stratified random audit). Unreviewed counts are in the table. Some source masks trace non-tumor structures (e.g. ventricles); confirmed cases are excluded, residual noise remains.

Limitations

One public dataset; 2-D slices; class balance and site distribution follow the source; not for clinical use.