- AI
- Molecular Imaging
- CT
- X-Ray
- Ultrasound
- MRI
- Facility Management
- Mammography
AI and Breast Ultrasound: Where Things Stand
What does the current evidence reveal for four FDA-cleared AI software modalities for breast ultrasound? Leading researchers review the study findings in a recently published literature review.
Can artificial intelligence (AI) reduce false-positive rates and unnecessary biopsies prompted by handheld and automated breast ultrasound modalities?
In a new literature review, recently published in the American Journal of Roentgenology, researchers provided their insights on current challenges with breast ultrasound, offered a closer look at the evidence for four FDA-cleared deep learning tools for adjunctive support and discussed future opportunities with the combination of AI and breast ultrasound.
Here are a few key takeaways from the literature review.
1. For breast lesion classification, the AI algorithm S-Detect (Samsung) had reported areas under curve (AUC) ranging from 83 to 89 percent and specificity rates ranging between 74 to 86 percent based on a meta-analysis and multicenter prospective study.
2. Use of the S-Detect software changed 87 out of 110 lesions (79 percent) from original BI-RADS 4A categories to BI-RADS 3 assessments, and only four of those lesions were cancerous, according to a multicenter prospective study, involving radiologists at secondary and rural hospitals who did not have expertise in breast ultrasound.
3. While multiple studies suggested that S-Detect could be beneficial for less experienced radiologists and one meta-analysis noted a pooled sensitivity rate of 82 percent, two multicenter prospective studies showed no difference in sensitivity between adjunctive use of S-Detect and unassisted radiologist interpretation.
A mass detected with supplemental ultrasound was categorized as a BI-RADS 4B lesion (A) but adjunctive use of the AI software S-Detect noted a possible malignancy. Subsequent surgical pathology showed the mass was invasive ductal carcinoma. (Images courtesy of the American Journal of Roentgenology.)
4. Two retrospective studies published in 2023 found that adjunctive use of the AI software Koios DS (Koios Medical) resulted in appropriate biopsy recommendations for 83 out of 83 cases of invasive lobular carcinomas and 333 out of 345 triple-negative cancers.
5. One multicenter study with 900 cancer-enriched breast lesions showed a four percent increase in mean reader AUC (83 to 87 percent) with adjunctive Koios DS. The study authors also noted class switching between benign and suspicions assessments in less than two percent of cases involving variations with region of interest (ROI) boundaries up to 20 percent.
6. In another study assessing the use of Koios DS for 200 breast lesions, the researchers noted no overall difference between adjunctive AI and unassisted reading. However, for cases in which radiologists expressed low confidence in lesion assessment, Koios DS demonstrated a six percent higher positive predictive value (25 percent vs. 19 percent) and a 13 percent higher specificity rate (58 percent vs. 45 percent).
7. Providing score of lesion (SLC) characteristics that correspond to BI-RADS categories, the AI software BU-CAD (TaiHao Medical) facilitated a seven percent AUC improvement over unassisted reading and reduced mean interpretation time from 30 to 18 seconds in a study involving 16 physicians assessing 172 cases. The researchers also pointed out that BU-CAD correctly diagnosed breast lesions in 163 out of 172 cases.
8. In one study involving six readers and 1,485 automated breast ultrasound (ABUS) images, the AI software QVCAD (QView Medical) demonstrated a 21 percent increase in sensitivity for readers with one to three years of experience in breast ultrasound (88 percent vs. 67 percent for unassisted reading).
9. In another study assessing QVCAD in 1,000 cases, researchers noted improved accuracy for inexperienced ABUS readers but no difference for experienced readers. Other research showed a seven percent AUC for all readers (84 percent vs. 77 percent) but lower specificity for four out of the six readers. In the same study, concurrent adjunctive use of QVCAD led to a 16 second reduction in mean reading time (34 seconds vs. 50 seconds for unassisted reading), according to the researchers.
The Reading Room: Artificial Intelligence: What RSNA 2020 Offered, and What 2021 Could Bring
December 5th 2020Nina Kottler, M.D., chief medical officer of AI at Radiology Partners, discusses, during RSNA 2020, what new developments the annual meeting provided about these technologies, sessions to access, and what to expect in the coming year.
Could a Deep Learning Model for Mammography Improve Prediction of DCIS and Invasive Breast Cancer?
April 15th 2024Artificial intelligence (AI) assessment of mammography images may significantly enhance the prediction of invasive breast cancer and ductal carcinoma in situ (DCIS) in women with breast cancer, according to new research presented at the Society for Breast Imaging (SBI) conference.
