Multimodal artificial intelligence for breast cancer detection in a population of women with dense breasts

Tao Tan, Alejandro Rodriguez Ruiz, Nico Karssemeijer, Ritse M. Mann, Lingyun Bao

Objective
To assess the combined performance of artificial intelligence ( detection systems for mammography and automated 3 D breast ultrasound) and their value to improve radiologists’ performance in detecting breast cancer in dense breasts

Method:

430 women with paired DM and ABUS examinations from an Asian population with predominantly dense breasts, with the following characteristics. Average age 48 years old (range 30 70), 42 w/ biopsy proven malignant lesion. 114 w/ benign lesions, 274 normal with 1 year follow up, 73 BI RADS breast density C or D. A subcohort of mammograms was selected for an observer study with 2 radiologists 7 and 10 years of experience) all 42 malignant cases and a total of 30 benign and 80 normal cases randomly selected.
AI for mammograms (AI DM) Transpara™ (v 1 5 0 ScreenPoint Medical) AI for automated 3 D breast ultrasound (AI ABUS) QVCAD 3 4 Qview Medical Inc Los Altos, California, USA
Based on automatically detected suspicious findings, each AI system assigns a continuous score to each examination representing the likelihood that cancer is present
Multi modal AI is computed by averaging normalized scores of the two single modal AI systems. Performances were compared using receiver operating characteristic analysis ( and their area under the curve. The tested hypotheses are AUC of multi modal AI is superior to single modal AI systems, AUC of the combination AI Radiologists is superior to the AUC of radiologists alone reading DM. The significance level of the analysis was 0 05 (P values and confidence intervals were adjusted for testing multiple hypotheses using Bonferroni correction) The iMRMC software (version 4 0 0 was used)

Results

Full cohort: AUC multi modal AI was higher (P=0.009) than AUC single modal AI.

Observer study cohort: Combination of radiologists scores with multi modal AI showed potential to improve radiologists’ performance (AUC AI multimodal Radiologists 0 853 0 050 AUC difference, P= 0 07

Conclusion
Automated multi modal ABUS+DM AI detection systems could be a potential solution to assist radiologists with breast cancer screening of women with dense breasts