File size: 8,136 Bytes
16971e4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 | # Scorer v2 report (rule-based, no LLM)
Scorer: `rule-v2` (scorer_v2.py + lexicon_v2.py + rubric.v2.json); self-test: scorer_v2_selftest.py (19 known cases, all must pass). Scope: every bank llava text in release r1 (TCGA 842 bank slides, BRACS final bank 460, EBR 2,019), 3 prompts each; BRACS scored twice (scope bracs-fine and bracs-coarse, same texts).
## Claims allowed
CONSISTENT / CONTRADICTED / UNINFORMATIVE means consistent or inconsistent with the confirmed label as defined by cited free sources (rubric.v2.json). It is NOT a statement of pathological correctness. No pathologist reviewed anything. The scorer rules and the earlier hand reads were written by the same author, and the 95.6% hand agreement is therefore weak support; the independent rater check is a v3 item.
**TCGA caveat: possibly train-contaminated (WSI-LLaVA was trained on TCGA; WSI-Bench is TCGA-derived). Every TCGA text-derived number below carries this caveat. The exact slide overlap with the WSI-Bench split was not checked.**
## Changes vs scorer v1 (KONGMING-REVIEW P0-3 / P0-4)
- Four stretched rubric entries repaired (secretory psammoma removed, secretory no-atypia and fibrous no-whorls downgraded, FEA columnar single-source).
- Rule fixes: normal-tissue phrase no longer matches normal-breast; negation across an adverb ("not prominently observed"); bare glioma is a hypernym compatible with any glioma label; hyperchromasia is not pleomorphism; hedges (may represent, unlikely, rather than); generic "invasive carcinoma" on lung is not ductal; N vs PB not separated by severity.
- Contradiction rates are reported by basis. Headline = diagnosis-driven only; feature-driven rows are listed apart.
## Flags per dataset / prompt
| scope | prompt | n | CONSISTENT | CONTRADICTED | UNINFORMATIVE | contradicted by diagnosis | stock-sentence share | texts identical to another |
|---|---|---|---|---|---|---|---|---|
| bracs-coarse | llava_limits | 460 | 24 | 33 | 403 | 0/460 = 0.0% (Wilson 95% CI 0.0-0.8, optimistic) | 89% | 97% |
| bracs-coarse | llava_morph | 460 | 145 | 144 | 171 | 97/460 = 21.1% (Wilson 95% CI 17.6-25.0, optimistic) | 60% | 72% |
| bracs-coarse | llava_type_grade | 460 | 156 | 303 | 1 | 303/460 = 65.9% (Wilson 95% CI 61.4-70.1, optimistic) | 79% | 68% |
| bracs-fine | llava_limits | 460 | 21 | 36 | 403 | 0/460 = 0.0% (Wilson 95% CI 0.0-0.8, optimistic) | 89% | 97% |
| bracs-fine | llava_morph | 460 | 104 | 198 | 158 | 99/460 = 21.5% (Wilson 95% CI 18.0-25.5, optimistic) | 60% | 72% |
| bracs-fine | llava_type_grade | 460 | 108 | 352 | 0 | 351/460 = 76.3% (Wilson 95% CI 72.2-80.0, optimistic) | 79% | 68% |
| ebr-coarse30 | llava_limits | 2019 | 1 | 0 | 2018 | 0/2019 = 0.0% (Wilson 95% CI 0.0-0.2, optimistic) | 92% | 97% |
| ebr-coarse30 | llava_morph | 2019 | 438 | 425 | 1156 | 245/2019 = 12.1% (Wilson 95% CI 10.8-13.6, optimistic) | 54% | 60% |
| ebr-coarse30 | llava_type_grade | 2019 | 407 | 1404 | 208 | 1387/2019 = 68.7% (Wilson 95% CI 66.6-70.7, optimistic) | 43% | 56% |
| tcga-nsclc | llava_limits | 842 | 2 | 0 | 840 | 0/842 = 0.0% (Wilson 95% CI 0.0-0.5, optimistic) | 91% | 98% |
| tcga-nsclc | llava_morph | 842 | 745 | 87 | 10 | 87/842 = 10.3% (Wilson 95% CI 8.5-12.6, optimistic) | 87% | 90% |
| tcga-nsclc | llava_type_grade | 842 | 648 | 194 | 0 | 194/842 = 23.0% (Wilson 95% CI 20.3-26.0, optimistic) | 64% | 68% |
Stock-sentence share = mean over texts of the fraction of sentences that (after lowercasing) appear in >=20 distinct slides of the same scope and prompt. "Identical" = whole text equal to another slide's text. Scope: bank slides only.
## Feature-driven contradictions (not for headline use)
| scope | prompt | feature-multi-source | feature-single-source |
|---|---|---|---|
| bracs-coarse | llava_limits | 0 | 33 |
| bracs-coarse | llava_morph | 0 | 47 |
| bracs-coarse | llava_type_grade | 0 | 0 |
| bracs-fine | llava_limits | 3 | 33 |
| bracs-fine | llava_morph | 19 | 80 |
| bracs-fine | llava_type_grade | 0 | 1 |
| ebr-coarse30 | llava_limits | 0 | 0 |
| ebr-coarse30 | llava_morph | 21 | 159 |
| ebr-coarse30 | llava_type_grade | 3 | 14 |
| tcga-nsclc | llava_limits | 0 | 0 |
| tcga-nsclc | llava_morph | 0 | 0 |
| tcga-nsclc | llava_type_grade | 0 | 0 |
## Per confirmed label (flags by prompt)
| scope | label | slides | morph C/X/U | type_grade C/X/U | limits C/X/U |
|---|---|---|---|---|---|
| bracs-coarse | atypical | 66 | 0/38/28 | 0/66/0 | 0/9/57 |
| bracs-coarse | benign | 233 | 0/105/128 | 0/232/1 | 0/24/209 |
| bracs-coarse | malignant | 161 | 145/1/15 | 156/5/0 | 24/0/137 |
| bracs-fine | ADH | 36 | 0/30/6 | 0/36/0 | 0/4/32 |
| bracs-fine | DCIS | 49 | 9/21/19 | 0/49/0 | 0/3/46 |
| bracs-fine | FEA | 30 | 0/13/17 | 0/30/0 | 0/5/25 |
| bracs-fine | IC | 112 | 95/1/16 | 108/4/0 | 21/0/91 |
| bracs-fine | N | 37 | 0/23/14 | 0/37/0 | 0/5/32 |
| bracs-fine | PB | 131 | 0/52/79 | 0/131/0 | 0/12/119 |
| bracs-fine | UDH | 65 | 0/58/7 | 0/65/0 | 0/7/58 |
| ebr-coarse30 | Adamantinomatous craniopharyngioma | 75 | 2/65/8 | 0/72/3 | 0/0/75 |
| ebr-coarse30 | Anaplastic astrocytoma, IDH-mutant | 37 | 2/12/23 | 2/35/0 | 0/0/37 |
| ebr-coarse30 | Anaplastic astrocytoma, IDH-wildtype | 37 | 2/6/29 | 6/31/0 | 0/0/37 |
| ebr-coarse30 | Anaplastic ependymoma | 40 | 0/2/38 | 0/16/24 | 0/0/40 |
| ebr-coarse30 | Anaplastic meningioma | 36 | 1/7/28 | 2/33/1 | 0/0/36 |
| ebr-coarse30 | Anaplastic oligodendroglioma, IDH-mutant and 1p/19q codeleted | 81 | 3/1/77 | 67/14/0 | 0/0/81 |
| ebr-coarse30 | Angiomatous meningioma | 21 | 0/5/16 | 0/19/2 | 0/0/21 |
| ebr-coarse30 | Atypical meningioma | 73 | 31/18/24 | 6/66/1 | 0/0/73 |
| ebr-coarse30 | Diffuse astrocytoma, IDH-mutant | 60 | 38/10/12 | 2/56/2 | 0/0/60 |
| ebr-coarse30 | Diffuse large B-cell lymphoma of the CNS | 49 | 13/3/33 | 21/26/2 | 0/0/49 |
| ebr-coarse30 | Ependymoma | 36 | 0/1/35 | 0/13/23 | 0/0/36 |
| ebr-coarse30 | Fibrous meningioma | 47 | 7/33/7 | 0/47/0 | 0/0/47 |
| ebr-coarse30 | Ganglioglioma | 78 | 1/4/73 | 37/11/30 | 0/0/78 |
| ebr-coarse30 | Glioblastoma, IDH-mutant | 24 | 16/0/8 | 2/14/8 | 0/0/24 |
| ebr-coarse30 | Glioblastoma, IDH-wildtype | 464 | 245/17/202 | 161/224/79 | 0/0/464 |
| ebr-coarse30 | Gliosarcoma | 49 | 32/1/16 | 0/45/4 | 0/0/49 |
| ebr-coarse30 | Haemangioblastoma | 78 | 6/10/62 | 1/75/2 | 0/0/78 |
| ebr-coarse30 | Haemangioma | 20 | 0/4/16 | 0/19/1 | 0/0/20 |
| ebr-coarse30 | Haemangiopericytoma | 24 | 0/2/22 | 0/23/1 | 0/0/24 |
| ebr-coarse30 | Langerhans cell histiocytosis | 22 | 0/2/20 | 0/22/0 | 0/0/22 |
| ebr-coarse30 | Lipoma | 28 | 0/17/11 | 0/28/0 | 0/0/28 |
| ebr-coarse30 | Medulloblastoma, non-WNT/non-SHH | 22 | 0/1/21 | 0/22/0 | 0/0/22 |
| ebr-coarse30 | Meningothelial meningioma | 94 | 0/64/30 | 0/93/1 | 0/0/94 |
| ebr-coarse30 | Metastatic tumours | 37 | 31/0/6 | 28/5/4 | 1/0/36 |
| ebr-coarse30 | Oligodendroglioma, IDH-mutant and 1p/19q codeleted | 75 | 2/0/73 | 72/3/0 | 0/0/75 |
| ebr-coarse30 | Pilocytic astrocytoma | 163 | 4/7/152 | 0/151/12 | 0/0/163 |
| ebr-coarse30 | Pituitary adenoma | 89 | 1/54/34 | 0/88/1 | 0/0/89 |
| ebr-coarse30 | Schwannoma | 71 | 1/10/60 | 0/68/3 | 0/0/71 |
| ebr-coarse30 | Secretory meningioma | 31 | 0/20/11 | 0/31/0 | 0/0/31 |
| ebr-coarse30 | Transitional meningioma | 58 | 0/49/9 | 0/54/4 | 0/0/58 |
| tcga-nsclc | LUAD | 427 | 418/6/3 | 305/122/0 | 2/0/425 |
| tcga-nsclc | LUSC | 415 | 327/81/7 | 343/72/0 | 0/0/415 |
C = CONSISTENT, X = CONTRADICTED, U = UNINFORMATIVE.
## BRACS non-IC framing
type_grade on BRACS non-IC bank slides: 347/348 contradicted by diagnosis; on IC slides: 4/112 contradicted. The rate mostly reflects the label mix (the model names a carcinoma on almost every slide), not a per-slide read.
## Use rules
- Cite a llava text only if CONSISTENT. Never cite CONTRADICTED or UNINFORMATIVE.
- llava_limits has no scorable content (boilerplate); do not say it carries no information.
- Wilson intervals assume independent slides; BRACS slides share patients (patient ids not in the release), EBR patient ids unknown: intervals are optimistic.
- Contradictions are label-space judgements ("label-inconsistent naming"), not morphological errors.
|