VIENNA -- Repurposed AI imaging tools can improve breast malignancy detection across diverse patient groups, suggest study findings presented February 27 at ECR 2025.
In her presentation, Helen Ngo from University Hospital Freiburg in Germany talked about her team’s research, which showed that this trend especially goes for mammograms of less dense breasts.
Helen Ngo presents her team's findings at ECR 2025, showing how repurposed AI tools can improve breast malignancy detection across diverse patient populations.
“AI is a reliable tool and shows promising diagnostic potential beyond screening,” Ngo said. “But careful interpretation is needed, especially in complex cases.”
While mammography interpretation by AI-based tools has been validated in screening settings, Ngo said that its role in diagnostic settings remains to be proven. This is especially true for symptomatic and post-treatment women.
Ngo and colleagues investigated the diagnostic performance of an AI tool (Lunit Insight MMG, Lunit) originally developed for screening mammography. The tool was repurposed for use in various clinical scenarios, including diagnostic mammograms in asymptomatic women, symptomatic women, and women with a personal history of breast cancer.
The study included data from 600 women collected between 2013 and 2023. Of the total diagnostic mammograms, asymptomatic women made up the majority (n = 410) followed by women with a personal history of breast cancer (n = 112), and then symptomatic women (n = 78). The AI tool scored the exams on a range from one to 100, with higher scores representing potential malignancy. The team used histopathological confirmation and/or follow-up of two years or more as the reference standard.
The AI tool achieved high marks in all study cohorts. Its performance further improved when the researchers excluded women with extremely dense breasts.
| Performance of breast AI tool in a variety of patient populations | |
|---|---|
| Patient subgroup | Area under the curve (AUC) |
| Asymptomatic women | 0.846 |
| Symptomatic women | 0.958 |
| Women with a personal history of breast cancer | 0.734 |
| Asymptomatic women (excluding dense breasts) | 0.874 |
| Symptomatic women (excluding dense breasts) | 0.96 |
| Women with a personal history of breast cancer (excluding dense breasts) | 0.827 |
When the team employed a threshold of the highest 10% AI scores to binarize the continuous AI output, the tool achieved a sensitivity of 92% and a specificity of 50% for asymptomatic women, 96% and 77% for symptomatic women, and 81% and 67% for women with a personal history of breast cancer, respectively.
Ngo said that future studies with larger and more diverse patient cohorts are needed to refine AI’s application. She added that this kind of research could also evaluate AI’s role in diagnostic breast imaging settings.
“Optimizing AI score interpretation, including threshold selection, requires larger datasets to ensure consistent performance across different clinical scenarios,” she said.
Read AuntMinnieEurope.com's entire coverage of ECR 2025 on our RADCast.
![Delineation workflow: The various steps of the contouring process: (a) the nipple is manually annotated (b) a reference point is placed 30 mm behind the nipple (c) 10 random theta/displacement values are selected to obtain the central point of each ROI (d) Regions of interest are placed on the image. (e) The resulting 10 ROIs obtained for a representative subject (viewing window [0 10000]). Images are published under a Creative Commons license (CC BY-NC-ND 4.0).](https://img.auntminnie.com/files/base/smg/all/image/2025/02/1_s2.0_S1120179725000523_gr1_lrg.67bcbbc3ac5a9.png?auto=format%2Ccompress&fit=crop&h=167&q=70&w=250)
