| """Labels the challenge's Gemma extractor reads back in addition to the one a bullet names. |
| |
| We ran 32 single-diagnosis reports through the extractor (google/gemma-4-31B-it, |
| temperature 0) and kept the extra labels that came back. Each link was checked |
| again inside a seven-bullet report; the one that did not hold there was dropped. |
| |
| A report is therefore read as the closure of its bullets under these links. |
| When the number of bullets is limited, bullets that imply other predicted labels |
| are preferred. |
| """ |
| import numpy as np |
|
|
| IMPLIES = { |
| "chronic_ischemic_stroke": ["encephalomalacia_gliosis"], |
| |
| |
| "intraparenchymal_hemorrhage": ["intracranial_hemorrhage"], |
| "glioma": ["brain_tumor"], |
| "brain_metastasis": ["brain_tumor"], |
| "meningioma": ["brain_tumor"], |
| "intraventricular_tumor": ["brain_tumor"], |
| "pituitary_tumor": ["brain_tumor"], |
| "hydrocephalus_ex_vacuo": ["cerebral_atrophy", "ventriculomegaly"], |
| "craniotomy_craniectomy": ["postsurgical_changes"], |
| "tumor_resection_cavity": ["postsurgical_changes"], |
| } |
|
|
|
|
| def closure_matrix(names): |
| """C[p, c] = 1 if asserting p makes the extractor also report c (transitively).""" |
| idx = {n: i for i, n in enumerate(names)} |
| L = len(names) |
| C = np.eye(L, dtype=bool) |
| for p, cs in IMPLIES.items(): |
| if p not in idx: |
| continue |
| for c in cs: |
| if c in idx: |
| C[idx[p], idx[c]] = True |
| for _ in range(L): |
| new = C | (C @ C) |
| if (new == C).all(): |
| break |
| C = new |
| return C |
|
|
|
|
| def apply_closure(asserted, C): |
| """asserted: bool[n, L] of what we write; returns what the extractor reads.""" |
| return (asserted.astype(np.uint8) @ C.astype(np.uint8)) > 0 |
|
|
|
|
| def select_with_implications(P, thr, cap, C): |
| """Greedy choice of at most `cap` bullets, valued over the implication closure. |
| |
| A label's value is its margin over its threshold, `P/thr - 1`. A bullet is worth |
| the total value of the labels it adds to the closure, so a parent whose children |
| are also predicted is chosen first, and a parent that would add unpredicted |
| children is penalised by their negative margin. |
| """ |
| n, L = P.shape |
| val = P / np.maximum(thr, 1e-9) - 1.0 |
| Cf = C.astype(np.float64) |
| if not cap: |
| return apply_closure(val > 0, C) |
| have = np.zeros((n, L), dtype=bool) |
| picked = np.zeros((n, L), dtype=bool) |
| rows = np.arange(n) |
| for _ in range(cap): |
| gain = (val * ~have) @ Cf.T |
| gain[picked] = -np.inf |
| d = gain.argmax(1) |
| g = gain[rows, d] |
| take = g > 1e-12 |
| picked[rows[take], d[take]] = True |
| have[take] |= C[d[take]] |
| if not take.any(): |
| break |
| return picked |
|
|
