Purpose: Deep learning has the potential to aid radiologists in image interpretation, improving diagnostic performance and patient care. The purpose of this study is to assess whether a developed deep learning model can aid radiologists at contrast enhanced spectral mammography (CESM) image classification.
Materials and Methods: This retrospective study included 108 women who underwent CESM between 2014-2019, with enhancing lesions on CESM. A total of 115 lesions were biopsied (69 benign, 46 malignant) and included in the analysis. Lesions were classified by a convolutional neural network (CNN) tool, and results were compared with the histopathology reports. Lesions were manually outlined by a radiologist, and BI-RADS features including lesion shape, outline and enhancement were filled-in by the same radiologist. Diagnostic performance of the CNN tool was assessed with and without the BI-RADS characteristics.
Results: Without the BI-RADS features, the CNN tool correctly identified 72/115 (62.6%) lesions. With the BI-RADS features, 79/115 (68.7%) lesions were correctly identified. Without the BI-RADS features the sensitivity of the tool was 95.7% (44/46) and specificity 40.6% (28/69). With the BI-RADS features sensitivity increased to 97.8% (45/46), and specificity to 49.3% (34/69). The false-negative case was an invasive lobular carcinoma.
Conclusion: A deep learning tool can aid radiologists in CESM interpretation, increasing specificity and reducing the number of false-positive cases.
