Anatomic location assignment rules

About this document

This is a near-verbatim migration of docs/anatomic-location-assignment-rules.md from the imaging-problem-list repository’s development branch.¹ The rules were agreed for a correction pass over the sample_data/example2 findings and are stated there as “the intended spec for tuning the automated location-coding step later.” Wording and structure follow the original; only terminology and links have been adjusted for this knowledgebase.

The subject is how an anatomicLocation of {locationId, locationDisplay}, a RadLex identifier from the anatomic-locations index, is assigned to each observation in an Exam Finding List.²

Precedence ladder

Apply per observation, in order.

1. Explicit anatomy in the report text or section context wins. Use the most specific structure stated, sided only if a side is stated. Section headers count as context. Under a Chest Wall: heading, “No soft tissue mass” resolves to chest wall; as a bare Soft Tissues: paragraph it stays the coarse region, per step 3.
2. Else, if the finding has a definable target organ or structure, use that organ at the finding’s own anatomic-noun granularity, never finer, applying the laterality rules below.
3. Else, meaning no anatomic noun at all, as in “soft tissue mass” or “generalized osteoporosis”, fall back to the exam-scoped coarse region (thorax, abdomen, head, and so on), sided only if the exam is sided.

Organ always wins over exam-region when the finding has a real target organ, especially in edge-of-exam cases. “No consolidation” or “lung bases clear” on an abdominal computed tomography resolves to lung, never “abdomen”.

Laterality

Source priority for laterality: explicit text side, then sided-exam side, then generic and unsided.

• A non-sided exam never introduces a side. On a computed tomography of the abdomen, “adrenal glands unremarkable” resolves to adrenal gland, generic, not “left adrenal gland”. “No hydronephrosis” with no side resolves to kidney, generic.
• A sided exam sides the finding. “XR Shoulder - left” plus “humerus fracture” resolves to left humerus.
• Resolve laterality against the full report section, not the finding’s isolated quote. A finding’s reportText is one verbatim snippet, and the side is often stated only in the section header or an earlier sentence of the same section, not in that snippet. When the section establishes a side, it applies to every finding in the section unless a later sentence contradicts it. This is precedence rule 1, text and section context wins, applied to laterality.

Worked example, a real miss now fixed. A chest computed tomography Shoulders: section reads “The left glenohumeral joint demonstrates a moderate joint effusion. The humeral head is in a high-riding position. Calcifications are noted in the region of the rotator cuff insertion.” The high-riding-head and calcific-tendinopathy findings carry snippets with no side, but the section establishes left, so they resolve to left glenohumeral joint and left supraspinatus tendon, not the generic structures.

Pitfall. Reviewing or coding from the isolated reportText silently drops section laterality and under-calls findings to the generic structure. Always read the snippet against its section.

Still do not infer a side across exams or from clinical priors. If this report’s section gives no side, stay generic even when another exam localized the same problem. An ultrasound snippet reading “parapelvic cysts” with no side stays kidney even though a prior computed tomography said “left kidney”.

Bilateral findings

• Positive bilateral, separable, meaning discrete independent lesions or structures per side, splits into two findings, one left and one right, each duplicating the original attributes and report text. Example: bilateral maxillary mucosal thickening becomes left and right maxillary sinus. Bilateral renal calculi were already separate observations in the source data and are not re-split.
• Positive bilateral, inseparable process, meaning a single diffuse entity named as one, an “-osis”, a “disease”, a confluent process, resolves to a generic, unsided structure. Example: bilateral cerebral small-vessel disease resolves to cerebral hemisphere, generic, not split.
• Spatially contiguous findings crossing the midline resolve to a single finding with an unsided structure.
• Generic absent, as in “No adrenal nodule” on a paired organ with no side stated, resolves to the generic organ, never a side.

Operational test for separable against inseparable: if you could meaningfully say “the left one is larger or newer”, it is separable; if it is a named diffuse entity, it is inseparable.

Specificity

The target’s granularity matches the finding’s anatomic noun, capped by the report, never finer.

• “humerus fracture”, from a cleared-bones statement, resolves to humerus, not “head of humerus”.
• “upper extremity fracture” resolves to upper extremity, a real body-part structure, sided by exam, not left blank.
• Region-named findings such as “upper extremity”, “skull”, and “spine” resolve to their body-part structure. They are step-2 targets, not the exam-region fallback.

Exam-scoped region fallback

Only for findings with no anatomic noun, such as “soft tissue mass” or “generalized osteoporosis”.

• The grain is the coarse region, not a soft-tissue substructure. “No soft tissue mass” on a chest computed tomography resolves to thorax, not “chest wall” or “musculature of thorax”, unless report section context names the substructure, in which case step 1 applies.
• Exam to region map, by RadLex identifier: computed tomography of the abdomen or abdomen and pelvis resolves to abdomen RID56; chest computed tomography or radiography to thorax RID1243; magnetic resonance of the brain to head RID9080; shoulder to shoulder RID39518. A combined abdomen and pelvis study defaults to abdomen unless the text points pelvic.
• Diffuse findings with a textual predominance keep the text location, per step 1. “Osteopenia, especially in the thoracic spine” resolves to thoracic vertebral column; a bare “Diffuse osteopenia” resolves to the exam region.

Resolution and validation

• For a named target organ, resolve by direct lookup, meaning AnatomicLocationIndex.get or an exact match, not flaky semantic search. This avoids retrieval misses such as “prostate” returning salivary glands. Derive sided identifiers from the generic entry through its left_id and right_id.
• Every assigned locationId must exist in the ontology. Structures absent from the ontology, such as basal cistern or dural venous sinus, are left unassigned rather than forced to a wrong code.

Known limitation

Parent-child or generic-versus-specific pairs for the same finding code across exams, such as “lung” against “lower lobe of right lung”, or “kidney” against “left kidney”, still produce separate Imaging Problem List groups. Deciding whether such observations are the same problem or distinct is the anatomic-compatibility reconciliation step, Step 3b of the anatomic location plan.³ That step has not been started.

Related

The anatomic locations area covers the underlying data model, the laterality triads these rules derive sided identifiers from, and the tooling. The pipeline that applies these rules is described in finding and location coding.

Footnotes

1. docs/anatomic-location-assignment-rules.md, imaging-problem-list dev branch ↩

2. The anatomic-locations package ↩

3. Anatomic location for Exam Finding Lists and Imaging Problem Lists, dev branch ↩