System Maintenance Notice

Please note that maintenance work is scheduled between 4 AM and 5 AM - CET. During this time, you may experience reduced performance. Thanks for your patience.

Publications

Journal Article Ahead of print

PubMed: 42183929 DOI: 10.1007/s11604-026-02016-3

Evaluating artificial intelligence-generated synthesized mammography as a standalone alternative to digital mammography with or without tomosynthesis.

Jpn J Radiol

Authors: Takayoshi Uematsu, Kazuaki Nakashima, Taiyo L Harada, Sachiko Yuen, Ichiro Isomoto, Hatsuko Nasu, Tatsuya Igarashi, Shoko Abe, Kouichirou Oosawa, Takashi Itoh, Yuji Maruo, Hidetaka Kobayashi, Yoshihiro Sakai

Affiliations

  • Division of Breast Imaging and Breast Interventional Radiology, Shizuoka Cancer Center Hospital, Shizuoka, 411-8777, Japan.
  • Division of Breast Imaging and Breast Interventional Radiology, Shizuoka Cancer Center Hospital, Shizuoka, 411-8777, Japan.
  • Department of Breast Surgery and Oncology, Shinko Hospital, Kobe, Japan.
  • Department of Radiology, St. Francis Hospital, Nagasaki, Japan.
  • Department of Radiology, Hamamatsu University School of Medicine, Shizuoka, Japan.
  • Department of Radiology, Nishiyaidu Health Check-Up Center, Shizuoka, Japan.
  • Department of Radiology, NHO Shizuoka Medical Center, Shizuoka, Japan.
  • Osawa Clinic, Shizuoka, Japan.
  • Division of Surgery, Seirei-Numazu Hospital, Shizuoka, Japan.
  • Onaribashi Sakae Clinic, Shizuoka, Japan.
  • Sakai Clinic, Shizuoka, Japan.

Abstract

PURPOSE

To evaluate whether standalone synthesized mammography (SM) can maintain or improve diagnostic accuracy while reducing reading time and radiation dose, compared to digital breast tomosynthesis (DBT) with digital mammography (DM) or DM alone.

MATERIALS AND METHODS

This was a retrospective study. SM images were generated using an AI-supported DBT reading method that synthesizes key features from DBT images without displaying marks or outlines. Twelve readers independently evaluated 200 breasts (including 50 cancers) using SM alone and DM with or without DBT acquired from October 2022 to January 2023. Diagnostic performance was assessed using area under the receiver operating characteristic curve (AUC), sensitivity, specificity, reading time, and radiation dose. A multireader, multicase statistical model was applied. Sample size and power calculations were based on estimates from a prior study.

RESULTS

SM alone achieved an AUC of 0.912, which did not differ significantly from that of DBT with DM (0.906; p = 0.538) or DM alone (0.897; p = 0.13). Sensitivity improved from 88.1% (DBT with DM) and 86.6% (DM) to 90.9% with SM, with a significant improvement over DM (p = 0.042). Specificity remained comparable (SM: 84.5%; DBT with DM: 85.9%; DM: 86.5%). SM alone reduced mean reading time by 53.1% (122.6 to 57.5 s; p < 0.001) and radiation dose by 39% (1.73 vs. 2.84 mGy) compared with DBT with DM.

CONCLUSIONS

Standalone AI-generated SM preserved the diagnostic accuracy of DBT with DM while reducing reading time and radiation dose, and improved sensitivity over DM without loss of specificity.

MeSH

Intelligence, Reading, Mammography, Radiology, Breast, Artificial Intelligence