Deep learning model with collage images for the segmentation of dedicated breast positron emission tomography images

Background: Dedicated breast positron emission tomography (dbPET) has high contrast and resolution optimized for detecting small breast cancers, leading to its noisy characteristics. This study evaluated the application of deep learning to the automatic segmentation of abnormal uptakes on dbPET to facilitate the assessment of lesions. To address data scarcity in model training, we used collage images composed of cropped abnormal uptakes and normal breasts for data augmentation. Methods: This retrospective study included 1598 examinations between April 2015 and August 2020. A U-Net-based model with an uptake shape classification head was trained using either the original or augmented dataset comprising collage images. The Dice score, which measures the pixel-wise agreement between a prediction and its ground truth, of the models was compared using the Wilcoxon signed-rank test. Moreover, the classification accuracies were evaluated. Results: After applying the exclusion criteria, 662 breasts were included; among these, 217 breasts had abnormal uptakes (mean age: 58 ± 14 years). Abnormal uptakes on the cranio-caudal and mediolateral maximum intensity projection images of 217 breasts were annotated and labeled as focus, mass, or non-mass. The inclusion of collage images into the original dataset yielded a Dice score of 0.884 and classification accuracy of 91.5%. Improvement in the Dice score was observed across all subgroups, and the score of images without breast cancer improved significantly from 0.750 to 0.834 (effect size: 0.76, P = 0.02). Conclusions: Deep learning can be applied for the automatic segmentation of dbPET, and collage images can improve model performance.