Datasets:

pengkaiScience
/

NPCBench

# MRI sequence-group data

This directory contains the group-format inputs for 300 T-stage and 400 N-stage patients. Together with the 100-patient structure-card panel in `L3/primary_l3_inputs/`, these form T400 and N400. Both tasks use AJCC/UICC edition 8; the T and N cohorts are patient-disjoint.

## Files

- `images-01.tar` through `images-07.tar`: seven 100-case image shards, containing 45,829 lossless PNG groups.
- `requests.jsonl`: prompts and ordered image paths for all 700 patients.
- `requests_gpt41.jsonl`: at most 50 groups per case, sampled within the axial/coronal groups with both endpoints retained.
- `requests_internvl.jsonl`: the groups selected for the model's 40,960-token context.
- `T300/labels.jsonl`, `N400/labels.jsonl`: report-derived reference stages.

Extract every shard into this directory. Image paths in these request files are relative to this directory. The complete T400 and N400 labels are one level above, in `L3/T400.jsonl` and `L3/N400.jsonl`.

## Layout and coverage

Axial groups stack T1, T2 and enhanced T1 from the same physical plane. Coronal groups stack enhanced fat-suppressed T1 and precontrast non-fat-suppressed T1. Each clinical display tile is 512×512 pixels. Group dimensions are 512×1644 and 512×1096 pixels, including sequence labels.

T coverage runs from the suprasellar cistern to inferior C3; N coverage extends from the skull base to the inferior sternoclavicular joint. The study team reviewed the anatomical bounds. All selected images for a patient enter one model request. Model-specific sampling, tiling and spatial merging are documented in the [evaluation code](https://github.com/pidneuralode/NPCBench/blob/main/docs/mri_extension.md).

The 300 group-format T cases and all 400 N cases have explicit clinical M0 status. The combined T400 set includes 374 M0, 14 M1 and 12 cases with unspecified M status. T and N labels are reference labels for their own tasks; the T300 set does not provide independent structure-invasion annotations.

Images are rendered derivatives without DICOM headers or patient-identity maps. The files contain anonymous case IDs and sequence labels. Clinical pixels are retained.

	# MRI sequence-group data

	This directory contains the group-format inputs for 300 T-stage and 400 N-stage patients. Together with the 100-patient structure-card panel in `L3/primary_l3_inputs/`, these form T400 and N400. Both tasks use AJCC/UICC edition 8; the T and N cohorts are patient-disjoint.

	## Files

	- `images-01.tar` through `images-07.tar`: seven 100-case image shards, containing 45,829 lossless PNG groups.
	- `requests.jsonl`: prompts and ordered image paths for all 700 patients.
	- `requests_gpt41.jsonl`: at most 50 groups per case, sampled within the axial/coronal groups with both endpoints retained.
	- `requests_internvl.jsonl`: the groups selected for the model's 40,960-token context.
	- `T300/labels.jsonl`, `N400/labels.jsonl`: report-derived reference stages.

	Extract every shard into this directory. Image paths in these request files are relative to this directory. The complete T400 and N400 labels are one level above, in `L3/T400.jsonl` and `L3/N400.jsonl`.

	## Layout and coverage

	Axial groups stack T1, T2 and enhanced T1 from the same physical plane. Coronal groups stack enhanced fat-suppressed T1 and precontrast non-fat-suppressed T1. Each clinical display tile is 512×512 pixels. Group dimensions are 512×1644 and 512×1096 pixels, including sequence labels.

	T coverage runs from the suprasellar cistern to inferior C3; N coverage extends from the skull base to the inferior sternoclavicular joint. The study team reviewed the anatomical bounds. All selected images for a patient enter one model request. Model-specific sampling, tiling and spatial merging are documented in the [evaluation code](https://github.com/pidneuralode/NPCBench/blob/main/docs/mri_extension.md).

	The 300 group-format T cases and all 400 N cases have explicit clinical M0 status. The combined T400 set includes 374 M0, 14 M1 and 12 cases with unspecified M status. T and N labels are reference labels for their own tasks; the T300 set does not provide independent structure-invasion annotations.

	Images are rendered derivatives without DICOM headers or patient-identity maps. The files contain anonymous case IDs and sequence labels. Clinical pixels are retained.