Restore L4 clinical chronology and add Full100 MRI schema condition
Browse files- L3/extensions/README.md +26 -0
- L4/questions.jsonl +0 -0
- README.md +13 -1
- supplementary/full100_schema/README.md +15 -0
- supplementary/full100_schema/mri_samples.jsonl +0 -0
L3/extensions/README.md
ADDED
|
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# T300 and N400 MRI extension
|
| 2 |
+
|
| 3 |
+
| Cohort | Cases | Reference-stage distribution |
|
| 4 |
+
|---|---:|---|
|
| 5 |
+
| T300 | 300 | T1: 35; T2: 35; T3: 115; T4: 115 |
|
| 6 |
+
| N400 | 400 | N0: 80; N1: 160; N2: 80; N3: 80 |
|
| 7 |
+
|
| 8 |
+
Both use AJCC/UICC edition 8 and explicitly documented cM0. T300 extends the
|
| 9 |
+
original T100; N400 is patient-disjoint from their combined T400 cohort. Original
|
| 10 |
+
T100 includes patients with cM1 or unspecified M status, so combined T400 is not
|
| 11 |
+
an all-M0 cohort. Original T100 results retain their original visual condition.
|
| 12 |
+
|
| 13 |
+
The extension uses five sequences: axial T1, axial T2, axial enhanced T1,
|
| 14 |
+
coronal enhanced fat-suppressed T1, and coronal precontrast non-fat-suppressed T1.
|
| 15 |
+
Matched axial and coronal planes form separate ordered image groups, jointly
|
| 16 |
+
supplied in each model request. A clinical display tile is 512 × 512 pixels.
|
| 17 |
+
|
| 18 |
+
Target axial coverage is suprasellar cistern to inferior C3 for T staging and
|
| 19 |
+
skull base to inferior sternoclavicular joint for N staging. The reviewed example
|
| 20 |
+
bounds are confirmed; patient-specific boundary review is not complete across
|
| 21 |
+
the extension. Image OCR screening is complete, with findings adjudication and
|
| 22 |
+
visual review still open. Consequently this directory currently contains the
|
| 23 |
+
cohort description only, not the 700-case image download or a final score table.
|
| 24 |
+
|
| 25 |
+
See the [evaluation guide](https://github.com/pidneuralode/NPCBench/blob/main/docs/mri_extension.md)
|
| 26 |
+
for the actual model-specific context, image encoding and sampling conditions.
|
L4/questions.jsonl
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
README.md
CHANGED
|
@@ -101,9 +101,21 @@ in cohort/node construction. Native CarePath questions use supplied clinical
|
|
| 101 |
context and do not recursively feed earlier model care answers back. Historical
|
| 102 |
25-patient controls use a different decision set.
|
| 103 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 104 |
## Ethics and use
|
| 105 |
|
| 106 |
-
|
| 107 |
|
| 108 |
The dataset contains authorized, de-identified benchmark derivatives under the
|
| 109 |
hospital-defined release scope and IRB B2026-465-01. It does not contain original
|
|
|
|
| 101 |
context and do not recursively feed earlier model care answers back. Historical
|
| 102 |
25-patient controls use a different decision set.
|
| 103 |
|
| 104 |
+
## Additional MRI evaluation
|
| 105 |
+
|
| 106 |
+
A separate [Full100 MRI prompt condition](supplementary/full100_schema/README.md)
|
| 107 |
+
provides the original Full25 JSON output schema for all 100 patients. The MRI
|
| 108 |
+
images, reference labels, pathology inputs and 217 scored care decisions are
|
| 109 |
+
unchanged. This condition is evaluated separately from the paper snapshot.
|
| 110 |
+
|
| 111 |
+
The [T300/N400 extension description](L3/extensions/README.md) summarizes the
|
| 112 |
+
new cohorts and input format. These additional image assets are not included in
|
| 113 |
+
the current download while clinical boundary review and image privacy review
|
| 114 |
+
are being completed. They are not included in the released counts above.
|
| 115 |
+
|
| 116 |
## Ethics and use
|
| 117 |
|
| 118 |
+
L4 retains clinical consultation, treatment and decision dates where they establish the applicable guideline edition. Birth-date expressions are removed; existing ages are retained, or an approximate age is stated with its reference year when only a birth date was available. Other longitudinal tasks use patient-relative days for converted complete dates. Answers, grouping and rubric weights are unchanged.
|
| 119 |
|
| 120 |
The dataset contains authorized, de-identified benchmark derivatives under the
|
| 121 |
hospital-defined release scope and IRB B2026-465-01. It does not contain original
|
supplementary/full100_schema/README.md
ADDED
|
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Full100 MRI prompt condition
|
| 2 |
+
|
| 3 |
+
`mri_samples.jsonl` contains 100 MRI requests using AJCC/UICC edition 9.
|
| 4 |
+
The 75 expanded-patient prompts now include the same JSON output example used
|
| 5 |
+
for the original 25 patients. The other 25 prompts are unchanged. All patient
|
| 6 |
+
IDs, images, labels and modality-source fields match the existing L6 download.
|
| 7 |
+
|
| 8 |
+
Image paths are relative to `L6/full100_inputs_scored217/`, obtained by extracting
|
| 9 |
+
`L6/carepath.tar.gz`. No duplicate image archive is needed. Keep the supplied
|
| 10 |
+
pathology proxy-image notices and the 217-care-node scoring policy.
|
| 11 |
+
|
| 12 |
+
Use `npcbench-infer --mri-prompts` with this file as described in the
|
| 13 |
+
[code guide](https://github.com/pidneuralode/NPCBench/blob/main/docs/mri_extension.md).
|
| 14 |
+
Care inputs must be rendered from the MRI and pathology predictions belonging
|
| 15 |
+
to this condition. This file supplies inputs, not completed model results.
|
supplementary/full100_schema/mri_samples.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|