Explainable artificial intelligence in medical ultrasound: a WoSCC-based bibliometric analysis, evidence map, and Scopus concordance assessment
TL;DR
Ultrasound XAI has expanded rapidly but growth in the literature has not been matched by comparable maturity in reported explanation evaluation or clinically relevant validation, and whether the leading bibliometric patterns are consistent across databases.
Abstract
Background Explainable artificial intelligence (XAI) is now used across medical ultrasound, yet it remains unclear whether the field has moved beyond producing explanation displays. This review examined whether the recent expansion of ultrasound XAI has been matched by explanation-specific evaluation and clinically relevant validation, and whether the leading bibliometric patterns are consistent across databases. Methods Web of Science Core Collection (WoSCC) and Scopus were searched for records published from 2010 to 2026 and indexed through July 9, 2026. After cross-database deduplication, 1,745 unique records were assessed and database-specific Broad and nested Core XAI datasets were constructed. WoSCC supported bibliometric mapping and structured evidence coding; Scopus and union datasets were used to assess incremental coverage and cross-database concordance. Results The searches retrieved 2,706 records and yielded 640 Union Broad and 514 Union Core XAI records; the primary WoSCC datasets contained 469 and 387 records, respectively. Publications rose sharply after 2023, with 2026 representing an incomplete year. SHAP (186/387, 48.1%) and CAM/Grad-CAM (79/387, 20.4%) predominated. An explanation output was reported or displayed without an identified evaluation procedure in 280/387 studies (72.4%), while no explicit explanation-evaluation information was identified in the reviewed evidence sources for 90/387 studies (23.3%); only 17 (4.4%) reported expert/reader, quantitative faithfulness or stability, or clinical-usefulness evaluation. External validation was reported in 62/387 records (16.0%), prospective design in 21/387 (5.4%), and multicenter design in 63/387 (16.3%). Country and cleaned-keyword ranks were highly concordant across databases (Spearman rho = 0.90 and 0.91), while Scopus added 36.5% Broad and 32.8% Core XAI records relative to WoSCC. Conclusion Ultrasound XAI has expanded rapidly, but growth in the literature has not been matched by comparable maturity in reported explanation evaluation or clinically relevant validation. Progress will require testable explanation claims and evaluation under realistic acquisition and clinical conditions.