Determine cohort equivalence for COLIPRI-CRM report evaluation

Determine whether the COLIPRI-CRM report-generation evaluation cohort is equivalent to either the 3,039-reconstruction or 1,564 report-level CT-RATE study protocol.

Background

The paper compares NV-Reason-CT with COLIPRI-CRM on CT-RATE report generation. COLIPRI-CRM is reported to achieve a RadBERT macro-F1 of 0.449, but the source study does not state how many images were included in its withheld test set. Because CT-RATE evaluations may operate either at the reconstruction level, using 3,039 reconstructed volumes, or at the acquisition/report level, using 1,564 studies, the comparability of the reported result cannot be established without identifying the COLIPRI-CRM cohort and its evaluation unit.

References

The authors report RadBERT macro-F1 0.449 on a withheld CT-RATE test set, but do not state its number of images. The cohort equivalence to either 3,039 reconstructions or 1,564 report-level studies remains unknown.

— NV-Reason-CT: 3D Visual Language Model for CT Analysis  (2609.27511 - Myronenko et al., 23 Sep 2026) in Section 4.1.2, CT-RATE report generation, paragraph “COLIPRI-CRM”