Dynamic AI-assisted ipsilateral tissue matching for digital breast tomosynthesis

To evaluate whether AI-assisted ipsilateral tissue matching in digital breast tomosynthesis (DBT) reduces localization errors beyond typical tumor boundaries, particularly for non-expert radiologists. The technology category is deep learning.

The study consisted of two parts. In Part 1, 14 radiologists subjectively evaluated the AI tool's impact on confidence and perceived usefulness on 11 cases. In Part 2, 12 radiologists marked lesion annotations with and without AI assistance after a washout period. Errors were measured relative to 94 consensus reference annotations in 30 cases established by two expert reviewers. Stratifications included expertise, lesion abnormality, and presence of ViewFinder warnings (indicating high uncertainty in the computed mapping). Wilcoxon signed-rank tests assessed statistical significance.

In Part 1, radiologists reported improved confidence and a mean usefulness score of 6.21/10 (p < 0.001), with greater benefit in challenging cases. In Part 2, root mean square error (RMSE) and maximum distance error (MDE) for abnormal lesions were 32 % (11.70 mm vs 8.88 mm) and 37.5 % higher (20.68 mm vs 15.08 mm), respectively, without AI assistance (p < 0.05). Non-expert readers showed the greatest benefit, with RMSE (12.20 mm vs 7.57 mm, p < 0.01) and MDE (15.76 mm vs 9.47 mm, p < 0.05) reduced by over 60 %. These reductions are clinically relevant given typical screen-detected breast tumor sizes (median, 13 mm [IQR: 9-20 mm]).

AI-assisted tissue matching significantly improved localization accuracy, especially for non- expert readers and in complex cases. The AI warning system effectively identified high-uncertainty cases. Errors were reduced to within typical tumor dimensions, potentially preventing missed lesions.