ICD-O morphology enum: completeness review against the knowledge base

Date: 2026-08-27 Scope: ICDOMorphologyEnum in src/dismech/schema/classifications/icdo_morphology.yaml, bound to classifications.icdo_morphology. Related: monarch-initiative/dismech#7548 (schema decision: four-digit ICD-O codes and a behaviour slot — still open, and not addressed here). Companion: cancer-taxonomy-granularity-review-2026-08-28.md covers the adjacent question — which cancer concepts get a Disease entry at all. It is about entry granularity; this report is about the morphology annotation those entries carry. The two are independent: an entry at any of its eight strata still needs a correct icdo_morphology value.

The finding

The vocabulary was not complete. All ten of its original values carried an ICD-O behaviour digit of /3, so it was a malignant-only histogenetic bucket wearing the name of the full ICD-O morphology axis. That had two consequences in the committed knowledge base:

Neoplastic entries with no correct value simply went unclassified, and four of them wrote the gap down in prose — Glomus Tumor (a CURATION_TODO discussion), Choriocarcinoma, Pheochromocytoma-Paraganglioma and GNAS-related pituitary adenoma 3. Those four notes are what this review acted on; they are the reason the omit-and-explain convention is worth keeping.
Entries with no correct value took a wrong one. Mesothelioma was tagged Carcinoma; polycythaemia vera, essential thrombocythaemia and primary myelofibrosis were tagged Leukemia; four germ cell entries were tagged Embryonal Neoplasm, an unrelated axis.

At review time 133 of ~280 neoplastic entries carried a value, and one value (Multiple Myeloma) was used by nothing.

The rule now written into the enum

A value names a morphology family, not an individual tumour entity. A family earns a value when the knowledge base holds entries no existing value can hold correctly and it is a top-level morphology group in ICD-O / the WHO classification. A sub-family is split out of its parent only when it dominates curation practice — which is why Adenocarcinoma and Squamous Cell Carcinoma sit beside Carcinoma, and Multiple Myeloma beside Plasma Cell Neoplasm, while single entities (glomus tumour, chordoma, GIST) are held by a family rather than given a value of their own.

Most new values are behaviour-neutral: Nerve Sheath Neoplasm covers schwannoma and MPNST alike. Where ICD-O itself splits a family on behaviour, the values follow it (Adenoma vs Adenocarcinoma). This is what unblocked Glomus Tumor: the objection to Sarcoma was never that the tumour is not mesenchymal, it was that Sarcoma asserts malignancy for an entity that is usually benign. A behaviour-neutral family value makes no such claim.

Values added (13)

Every meaning was checked against the NCI EVS REST API on 2026-08-27 for active status, preferred label and ICD-O-3_Code; every exact_mappings MONDO term was checked against OLS4 for non-obsolescence. An ICDO: mapping is recorded only where the bound NCIT concept itself carries an ICD-O-3_Code annotation, rather than hand-typing a code.

Value	NCIT	ICD-O-3	KB entries it unblocks
`Adenoma`	C2855	8140/0	adrenal cortical adenoma, the five pituitary adenoma entries, mucinous cystadenoma, the adenomatous polyposis entries
`Trophoblastic Tumor`	C3422	—	choriocarcinoma, gestational trophoblastic neoplasm
`Mesothelial Neoplasm`	C3786	—	pleural and peritoneal mesothelioma
`Pericytic Neoplasm`	C6528	—	glomus tumour
`Nerve Sheath Neoplasm`	C4972	—	schwannoma, neurofibroma, granular cell tumour, MPNST, the neurofibromatosis entries
`Meningioma`	C3230	9530/0	meningioma
`Germ Cell Tumor`	C3708	—	seminoma, embryonal carcinoma, yolk sac tumour, mixed and ovarian/testicular/CNS germ cell tumours
`Sex Cord-Stromal Tumor`	C3794	8590/1	granulosa cell tumour, Sertoli-Leydig cell tumour, testicular sex cord-stromal neoplasm
`Neuroendocrine Neoplasm`	C3809	—	phaeochromocytoma-paraganglioma, pancreatic and GEP NETs, large-cell and small-cell neuroendocrine carcinoma
`Plasma Cell Neoplasm`	C4665	—	plasma cell neoplasm (family), AL amyloidosis, POEMS
`Myeloproliferative Neoplasm`	C4345	9960/3	polycythaemia vera, essential thrombocythaemia, primary myelofibrosis, MPN-U
`Myelodysplastic Syndrome`	C3247	9989/3	myelodysplastic syndrome
`Histiocytic and Dendritic Cell Neoplasm`	C9294	—	Langerhans cell histiocytosis, Rosai-Dorfman disease

One correction to the research recorded on #7548: NCIT:C3234 (Mesothelioma) is a retired concept (conceptStatus = Retired_Concept, active = false, parented under "Retired Concept 2023"). Mesothelial Neoplasm is bound to NCIT:C3786 instead, which is active and sits directly under NCIT's Neoplasm by Morphology axis.

That retirement propagates into MONDO, and it is worth stating because it caught this review out. MONDO:0005065 "mesothelioma" is the obvious-looking mapping and is the wrong one twice over: it still xrefs the retired NCIT:C3234, and its definition — "a usually malignant and aggressive neoplasm of the mesothelium" — is behaviour-leaning, which is precisely what a behaviour-neutral family value must not assert. The correct term is MONDO:0006856 "mesothelial neoplasm", which xrefs NCIT:C3786 and is defined as "a benign or malignant neoplasm arising from mesothelial cells".

The general check, which is cheap and worth running on any future addition: the MONDO term should carry a database_cross_reference to the same NCIT code the value's meaning binds. All thirteen values here satisfy it.

No `Other` value

7548 suggested adding one. This review deliberately did not. The four prose

notes that drove this expansion only exist because curators had nowhere to put a value and wrote down why; an Other bucket would have absorbed exactly those signals and left the vocabulary frozen. The convention is now stated in the enum description and the disease-classification skill: when no value fits, omit the slot and record why in notes or a CURATION_TODO discussion.

Entries changed (31)

Nine corrections — values that asserted the wrong histogenesis:

Entry	Was	Now
`Malignant_Mesothelioma`	Carcinoma	Mesothelial Neoplasm
`Choriocarcinoma`	Carcinoma	Trophoblastic Tumor
`Polycythemia_Vera`	Leukemia	Myeloproliferative Neoplasm
`Essential_Thrombocythemia`	Leukemia	Myeloproliferative Neoplasm
`Primary_Myelofibrosis`	Leukemia	Myeloproliferative Neoplasm
`Testicular_Seminoma`	Embryonal Neoplasm	Germ Cell Tumor
`Embryonal_Carcinoma`	Embryonal Neoplasm	Germ Cell Tumor
`Mixed_Germ_Cell_Tumor`	Embryonal Neoplasm	Germ Cell Tumor
`Malignant_Germ_Cell_Tumor_of_Ovary`	Embryonal Neoplasm	Germ Cell Tumor

The germ cell group is the sharpest case. NCIT:C3752 (Embryonal Carcinoma) has parents Malignant Germ Cell Tumor and Nongerminomatous Germ Cell Tumor, both under NCIT:C3708; the children of NCIT:C3264 (Embryonal Neoplasm) are the blastomas, the CNS embryonal tumours, Ewing sarcoma/pPNET, rhabdoid tumour and Wilms tumour. The old value was a name collision, not a classification.

Twenty-one exemplars backfilled to exercise each new value, at least one per value, each carrying a notes line recording the NCIT ancestry it rests on: adrenal cortex adenoma, GNAS-related pituitary adenoma 3, gestational trophoblastic neoplasm, malignant peritoneal mesothelioma, glomus tumour, schwannoma, neurofibroma, granular cell tumour, meningioma, yolk sac tumour, testicular germ cell tumour, adult granulosa cell tumour of ovary, testicular sex cord-stromal neoplasm, pheochromocytoma-paraganglioma, pancreatic neuroendocrine tumour, plasma cell neoplasm, multiple myeloma, myeloproliferative neoplasm unclassifiable, myelodysplastic syndrome, Langerhans cell histiocytosis, Rosai-Dorfman disease.

And one exemplar withdrawn. Hydatidiform_Mole was assigned Trophoblastic Tumor in an earlier draft of this PR and the assignment has been removed, because it failed the same rule every other value passes. Hydatidiform mole does carry ICD-O morphology codes — 9100/0 for the mole NOS, 9103/0 for the partial mole — which is what made it look assignable. But NCIT:C3110 (Hydatidiform Mole) does not descend from NCIT:C3422 (Trophoblastic Tumor): its parents are Gestational Trophoblastic Disorder and Placenta Disorder, NCIT asserts no Neoplastic_Status for it, and both NCIT:C4871 (Complete) and NCIT:C4293 (Partial Hydatidiform Mole) sit under Placental Non-Neoplastic Disorder. WHO agrees — gestational trophoblastic neoplasia is invasive mole, choriocarcinoma, PSTT and ETT, and a non-invasive mole is GTD but not GTN. This axis applies only to neoplastic diseases, so the entry now omits the slot and records why, which is the convention this review argues for. The molar pregnancies that do progress are covered by Gestational_Trophoblastic_Neoplasm, which carries the value correctly.

Every one of the 23 values is now used by at least one entry — Multiple Myeloma, unused since the enum was written, went to the Multiple_Myeloma entry. The KB now carries 158 icdo_morphology assignments in total. (Note that the raw before/after pair is not a clean measure of this PR: the branch was rebased mid-review and picked up neoplastic entries added on main in the meantime, so the denominator moved underneath it.)

The prose notes on Glomus Tumor, Choriocarcinoma, Pheochromocytoma-Paraganglioma and GNAS-related pituitary adenoma 3 were rewritten to say what is now assignable and what is still blocked. The Glomus Tumor CURATION_TODO discussion stays open: the morphology half is solved, the four-digit codes (8711/0 benign; 8710/3 and 8711/3 for glomangiosarcoma, both on NCIT:C4221) and the behaviour digit are not.

What is still not covered

These are families that remain without a home. None passed the ≥1-KB-entry and top-level-ICD-O-group test cleanly enough to add in this pass, but they are the next candidates if more entries accumulate:

Family	KB entries	Note
Thymic epithelial neoplasms	`Thymoma`, `Thymus_Neoplasm`	NCIT:C3411 Thymoma is ICD-O 8580/1, under Thymus Epithelial Neoplasm; `Carcinoma` asserts both malignancy and the wrong histology
Gastrointestinal stromal tumour	`Gastrointestinal_Stromal_Tumor`	NCIT:C3868, ICD-O 8936/1, parent Stromal Neoplasm. Left at `Sarcoma`, which overstates behaviour. A single entity, not a family — the right fix is the behaviour slot from #7548, not a new value
Mast cell neoplasms	`Systemic_Mastocytosis`, `Maculopapular_Cutaneous_Mastocytosis`	NCIT parents Mastocytosis → Mast Cell Neoplasm; not histiocytic/dendritic, and WHO no longer files mastocytosis under MPN
Fibroblastic/myofibroblastic	`Solitary_Fibrous_Tumor`, `Inflammatory_Myofibroblastic_Tumor`, `Infantile_Myofibromatosis`	NCIT Fibroblastic Neoplasm; ICD-O 8815/0, 8815/1 and 9051/0 all on C7634
Perivascular epithelioid cell (PEComa)	`Perivascular_Epithelioid_Cell_Neoplasm`, `Lymphangioleiomyomatosis`	NCIT files these under Soft Tissue Neoplasm of Uncertain Differentiation, not under Pericytic Neoplasm — do not reach for the new value here
Odontogenic	`Ameloblastoma`
Craniopharyngioma	`Craniopharyngioma`	NCIT:C2964, ICD-O 9350/1, parents include Benign Squamous Cell Neoplasm — `Squamous Cell Carcinoma` would be wrong twice over
Benign melanocytic	`Nevus_of_Ota`, `Meningeal_Melanocytoma`	`Melanoma` is malignant-only
Neuronal and mixed neuronal-glial	`Dysembryoplastic_Neuroepithelial_Tumor`, `Mixed_Neuronal-Glial_Tumor`	not purely glial
Benign mesenchymal / fibroepithelial	`Uterine_Leiomyoma`, `Breast_Fibroadenoma`, `Bone_Giant_Cell_Tumor`	`Sarcoma` asserts malignancy

One entry that looks uncovered is not: Chordoma is assignable as Embryonal Neoplasm today, because NCIT:C2947 (Chordoma) sits under Notochordal Tumor, a direct child of NCIT:C3264. It is left unassigned pending a curator who is comfortable with that placement rather than assigned on the strength of the ontology path alone.

Separately, roughly 120 neoplastic entries carry no icdo_morphology despite an existing value fitting them (colon adenocarcinoma, Hodgkin lymphoma, osteosarcoma, angiosarcoma, hairy cell leukaemia, MALT lymphoma, and so on). That is a curation backlog, not a vocabulary gap, and is out of scope here.

A note for the next enum extension

A change like this one needs the scope-override label on its PR. CI's curation-scope guard fails any PR that touches kb/ and anything outside kb/, stubs/, history/, cache/, references_cache/, research/, pages/ or dashboard/ — and an enum extension is inherently both halves at once: the schema value and the entries that exercise it. Splitting them is worse, not better, because the KB half cannot validate until the schema half has merged. The guard exists to catch curation PRs that pick up infra changes by accident; this is the intentional case it provides the label for.

What this review did not do

Everything in #7548 that needs a schema decision rather than a vocabulary extension is untouched and still assigned to @cmungall:

DiseaseMappings.icdo_mappings for the four-digit NNNN/N code (multivalued — NCIT:C4221 carries two).
A first-class neoplastic_behavior slot for the ICD-O /0–/3 digit.
Ingesting NCIT:P334 (ICD-O-3_Code) and NCIT:P363 (Neoplastic_Status) through OntologyEdgeSource so those assertions become snippet-validated evidence.

The generated src/dismech/datamodel/dismech_pydantic.py was not regenerated: it is already ~12,500 diff lines stale against main's schema, so regenerating it here would bury this change. Nothing in the repository imports it and no test checks it, but it is worth a dedicated refresh PR.