Can AI-Powered Virtual Biopsies Improve Detection of Breast Lesion Subtypes on Digital Mammography?

New bi-national research suggests that artificial intelligence (AI) can help identify breast lesion subtypes on mammography images and potentially reduce unnecessary biopsies.

In a recently published retrospective study in radiologyThe researchers examined AI-enabled algorithms (trained using digital mammography and clinical data) in two different cohorts of women who had a one-year clinical and imaging history prior to biopsy diagnosis. The training sets included 2,120 women from the Israeli population (mean age 51) and 1,642 women from the US population (mean age 61), according to the study. The researchers found that testing sets consisted of 441 women (mean age 51) for the Israeli AI model and 344 women (mean age 60) for the US AI model.

When imaging and clinical information were integrated into the AI ​​models, the researchers found that the US AI model had an area under the curve (AUC) of 83 percent for detecting invasive cancer, an AUC of 74 percent for had ductal carcinoma in situ and an AUC of 72 percent for benign lesions. The Israeli AI model had an AUC of 85 percent for invasive carcinoma, an AUC of 76 percent for ductal carcinoma in situ, and an AUC of 82 percent for benign lesions, according to the study.

The researchers adjusted the prediction threshold of the AI ​​models to the radiological assessment of breast lesions via the Breast Imaging Reporting and Data System (BIRADS) and provided subsequent sensitivity rates of 98.7 percent and 96.8 percent for malignant lesions in the respective test sets for the Israeli AI fixed model and the US AI model. In assessing the impact of the AI ​​model’s sensitivity to false positives, the study authors suggested that the Israeli AI model, with a sensitivity of 99 percent, may have enabled a 13 percent reduction in unnecessary biopsies.

“Our preliminary results are encouraging and show that our models have the potential to reduce biopsy sample errors and lower costs associated with false positives,” wrote Michal Rosen-Zvi, Ph.D., the director of AI for Accelerated Healthcare and Life Sciences Discovery at IBM Research and Professor at the Hebrew University of Jerusalem in Israel and colleagues. “By grouping the pathological findings of breast lesions according to similar clinical treatment guidelines, our model offers the potential to help radiologists recommend the appropriate follow-up approach based on the assessment of clinical, radiological, and pathological considerations.”

(Editor’s note: For related content, see Mammography News: FDA Says National Breast Density Notification Rule May Be Published in Late 2022 or Early 2023, Digital Breast Tomosynthesis and Breast Density: What a New Study Reveals, and What a New meta-analysis reveals breast density, mammography and MRI screening.”

Rosen-Zvi and colleagues found that clinical data did not significantly affect the ability of the AI ​​models to detect benign lesions, ductal carcinoma in situ, or invasive carcinoma. The Israeli AI model had an 88 percent AUC for detecting malignant lesions, but the researchers found that imaging alone in the AI ​​model achieved an 87 percent AUC. They also pointed out that including clinical data in the US AI model did not affect cancer prediction.

Regarding the limitations of the study, the authors acknowledged a lack of generalizability of the results of the AI ​​model in different populations. Rosen-Zvi and colleagues said that testing the Israeli AI model in the US study population resulted in reduced differentiation between malignant and benign lesions, and testing the US AI model in the Israeli study population reduced detection of invasive carcinoma , ductal carcinoma in situ and general prediction of malignant lesions.

They suggested that incomplete data collection and the diversity of digital mammography workstations may have contributed to the lower performance of the US AI model across different hospital systems. Noting that the Israeli population is predominantly Jewish and the US study population had a small number of Jewish women (less than 1.1 percent), the study authors said differences in ethnic composition played a role in training the AI ​​algorithms could have different outcomes in populations.