Type:
Educational Exhibit
Keywords:
Artificial Intelligence, Mammography, Computer Applications-3D, Multidisciplinary cancer care
Authors:
S. Morrell, J. Cardoso, B. Batohi, K. Satchithananda, R. K. Wasan, R. Rahim, S. Ourselin, E. F. Conant, J. C. Morel
DOI:
10.26044/ecr2026/C-23895
Background
BACKGROUND
- Why localisation is hard in DBT: CC and MLO views are non‑orthogonal, tissue deforms between compressions and changes appearance, and each view contains ~60–80 depth slices; the reader must mentally map 3D tissue between views while scrolling. This increases cognitive load, fatigue and “false correspondence” (matching to the wrong tissue). [1,5–7]
- Why it matters: incorrect cross‑view matching can lead to mis‑targeted ultrasound (US)/biopsy (false reassurance), unnecessary additional imaging (repeat targeted views or MRI), and reduced confidence/efficiency in assessment workflows. [1,2]
- Why interval cancers matter: Interval cancers represent a substantial minority of breast cancers diagnosed in screened populations (often reported ~20–30%). They are an important marker of the effectiveness of the breast screening program; the lower the rates of interval cancers, the more effective the screening is. On retrospective review they are commonly grouped as (i) occult/true interval (not visible on the prior screen), (ii) subtle/minimal‑sign, and (iii) missed at the prior screen. Human factors contribute across categories (ii) and (iii), especially search/localisation, recognition, and decision thresholds, and DBT’s larger search volume and reader fatigue may amplify these mechanisms.
- Motivation: a pilot annotation study reported a 23% error rate in manual tissue matching even among experienced readers, and institutional exemplars show that plausible “looks‑similar” mismatches can delay diagnosis illustrate real world risk. [1]
- Rationale for AI assistance: an interactive, regulator‑cleared localisation aid (FDA 510(k) K223501) provides dynamic ipsilateral tissue matching between CC↔MLO with an uncertainty warning when mapping confidence is low; it does not classify lesions and does not replace radiologist judgement. [1]
- Localisation Communication: being able to give instructions and locality for subsequent procedures is beneficial for less experienced sonographers, sonographers who are not trained in mammography interpretation and others.
PROCEDURE DETAILS
A) How dynamic tissue matching works in practice
- “Point‑and‑map”:moving the cursor in one DBT view auto‑scrolls the paired view to the predicted corresponding slice and displays a confidence‑scaled ellipse around the likely corresponding tissue.
- Coronal/equidistance references: coronal views (compressed/uncompressed) and nipple–pectoralis equidistance guides can be used to sanity‑check geometry and support communication.
- Uncertainty management: a warning indicates high‑uncertainty mappings (e.g., certain physiology, atypical positioning/compression). Treat warnings as a prompt to slow down, verify positioning, and revert to manual multi‑slice matching. [1]
B) Evidence snapshot
A two‑part multi‑site study evaluated the impact of an AI tissue‑matching tool in assessment DBT cases (12 radiologists; 30 cases; 94 expert‑consensus reference annotations). Subjectively, radiologists reported increased confidence and a mean usefulness score of 6.21/10 (p≤0.001), with higher usefulness in challenging cases. Objectively, for abnormal regions of interest (ROIs) without warnings, localisation errors were significantly larger without AI assistance: mean root mean square error (RMSE) 70 mm (AI OFF) vs 8.88 mm (AI ON)and mean maximum distance error (MDE) 20.68 mm vs 15.08 mm(both p<0.05). Non‑expert readers showed the largest improvements (RMSE 12.20→7.57 mm, p<0.01; MDE 15.76→9.47 mm, p<0.05). These reductions are clinically relevant because the median screen‑detected tumour diameter is ~13 mm and T1 tumours are ≤20 mm; large mislocalisation risks interrogating normal tissue beyond lesion margins. The warning system identifies high‑uncertainty cases where manual correlation is preferred. [1]