OBJECTIVE To determine candidate lesion-based criteria for a positive sacroiliac joint (SIJ) MRI based on bone marrow oedema (BMO) and/or erosion in non-radiographic axial spondyloarthritis (nr-axSpA); to compare the performance of lesion-based criteria with global evaluation by expert readers. METHODS Two independent cohorts A/B of 69/88 consecutive patients with back pain aged ≤50 years, with median symptom duration 1.3/10.0 years, were referred for suspected SpA (A) or acute anterior uveitis plus back pain (B). Patients were classified according to rheumatologist expert opinion based on clinical examination, pelvic radiography and laboratory values as having nr-axSpA (n=51), ankylosing spondylitis (n=34) or non-specific back pain (n=72). Four blinded readers assessed SIJ MRI, recording the presence/absence of SpA by concomitant global evaluation of T1-weighted spin echo (T1SE) and short τ inversion recovery (STIR) scans and, thereafter, whether BMO and/or erosion were present/absent in each SIJ quadrant of each MRI slice. We derived candidate lesion-based criteria based on the number of SIJ quadrants with BMO and/or erosion and calculated mean sensitivity and specificity for SpA. RESULTS For both cohorts A/B, global assessment showed high specificity (0.95/0.83) compared with the Assessment in SpondyloArthritis international Society (ASAS) definition (0.76/0.74). BMO ≥3 (0.89/0.84) or ≥4 (0.92/0.87) showed comparably high specificity to global assessment. Erosion ≥2 and/or BMO ≥3 or ≥4 were associated with comparably high sensitivity to global assessment without affecting specificity. These combined criteria showed both higher sensitivity and specificity than the ASAS definition. CONCLUSIONS Lesion-based criteria for a positive SIJ MRI based on both BMO and/or erosion performed best for classification of axial SpA, reflecting the contextual information provided by T1SE and STIR sequences.