C) Four cases: failure modes, imaging cues, and how to improve
Case 1: Preventing a miss from a plausible manual mismatch: • Pitfall:“looks‑similar” posterior densities are manually paired across views, resulting in a false correspondence and US performed at the wrong site.• How AI helps: mapped tissue directs attention to the correct paired‑view location at index assessment (and flags uncertainty when geometry is unreliable).• Take‑home: treat easy appearance‑based matches with scepticism; triangulate using mapping + coronal/equidistance checks. [1]
Case 2: Efficient US targeting when the correlate is occult in one view: • Pitfall: an asymmetry is conspicuous in one view but the correlate is not obvious in the other, prompting broad US searching.• How AI helps: mapping predicts a specific quadrant/clock‑face location, enabling focused US to confirm a benign correlate and possibly avoid a low yield MRI.• Take‑home: start US at the mapped clock‑face and expand only if needed; document if a warning is present.
Case 3: Asymmetry / architectural distortion with normal ultrasound: avoiding unnecessary same-day work-up
• Pitfall: an apparent asymmetry / architectural distortion (AD) is visible on the CC view, but without reliable cross-view localisation information the expected MLO location is uncertain, prompting broad or precautionary same-day ultrasound; uncertainty can drive “problem‑solving” MRI after normal US..• How AI helps: tissue-matching software predicts the expected corresponding MLO location, narrowing the search area and supporting confident cross-view correlation when the finding is benign.• Take-home: when mapping and coronal/equidistance references are concordant and no warning is present, radiologists may be able to rule out the asymmetry more efficiently and, protocols permitting, avoid unnecessary ultrasound. [1,2]
Case 4: Ill‑defined asymmetry: surfacing subtle distortion• Pitfall: satisfaction‑of‑search and slice burden can cause subtle distortion within an asymmetry to be missed.• How AI helps: mapping highlights a focal area within the broader asymmetry; targeted US demonstrates subtle distortion with malignant pathology.• Take‑home: use mapping to “re‑interrogate” subtle regions, especially when the initial US is negative/equivocal. [5–7]
D) Practical localisation checklist (workstation + US communication)
- Start where the finding is clearest; note slice number and depth (anterior/mid/posterior).
- Cross‑view verify using three cues: (i) dynamic mapping/ellipse, (ii) coronal views, (iii) nipple–pectoralis equidistance references (when available).
- If a warning appears: verify positioning (nipple in profile, MLO angle, compression artefacts) and perform careful manual multi‑slice matching; do not “force” a correspondence.
- Communicate to US: breast side, quadrant + clock‑face, depth, and expected slice range; state uncertainty if warnings/positioning issues are present.
- If targeted US is negative: re‑check whether the correct mapped region was scanned before concluding “no correlate”. [1,2]
E) Safety, limitations, and future directions• AI matching is a localisation aid only; it does not provide malignancy classification and should not override imaging features or clinical judgement. [1]• Use is not validated in non‑standard positioning or with implants/major postsurgical change; heed warnings and fall back to manual correlation. [1]• This educational poster illustrates selected cases. The published evidence is a spatial‑accuracy reader study; diagnostic outcomes (cancer detection rate, recall, interval cancer reduction) were not measured. Future prospective studies should test whether improved localisation translates into fewer mis‑targeted work‑ups and fewer missed cancers (including interval cancers). [1,3,4]• Potential system implications (hypothesis‑generating): more accurate cross‑view localisation may reduce repeat/extended US searching and some low‑yield “problem‑solving” MRI; formal outcomes and health‑economic evaluations are ongoing.• Disclosure: some authors report financial relationships with the software developer; see manuscript declarations. [1]