Author: Xiaoyu Duan, Xiang Li, Wenbo Wan, Lei Zhang, Yiwen Zhang 👨🔬
Affiliation: Duke University, Medical Physics Graduate Program, Duke Kunshan University 🌍
Purpose: Breast screening has been proved to reduce breast cancer mortality by early detection and treatment for patients. Mammography is the most common and widely used technique for breast cancer screening. However, the high volume of images to be reviewed often results in high workloads, fatigue and increased risk errors of radiologists. This study aims to develop a deep-learning convolutional neural network-based breast cancer localization model to enhance the efficiency and diagnostic accuracy of radiologists during image reading process.
Methods: 960 mammographic images of digital anthropomorphic breast phantoms with lesion inserted were simulated via raytracing and split into 700, 80, and 180 for training, validation and testing, respectively. U-Net models were built with varying depths (3-6 layers) and different loss functions including Binary Cross Entropy (BCE), Dice-Sørensen Coefficient (Dice) and BCE with Dice. The U-Net model was initially pretrained on simulated dataset, then fine-tuned with transfer learning for clinical images from CBIS-DDSM dataset. The Euclidean Distance (ED) and Hausdorff Distance (HD) metrics evaluated for localization accuracy.
Results: Among different layers configurations, 6-layers achieved the best overall performance. Regarding loss functions, BCE with Dice had the highest overall scores. The model with 6 layers and BCE with Dice loss function showed highest accuracy for simulated images with mean (± standard deviation) values of ED (1.75 pixels ± 1.36), HD (12.93 pixels ± 20.42), SSIM (0.98 ± 0.32) and Dice (0.80 ± 0.07). Clinical images also presented accurate lesion localization results, with mean values of ED (7.76 pixels) and HD (40.07 pixels) for four testing images.
Conclusion: This study highlights the potential of U-Net for breast cancer lesion localization, achieving high accuracy on simulated data. It also shows promising prospects for application to clinical images. Future research includes refining the model architecture for robust and accurate application on larger patient cohort.