MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Mass Detection and Segmentation in Digital Breast Tomosynthesis Via Deep-Learning

G Qin1,2*, H Chen3 , H Zeng2 , Y Xu3 , Z Zhou1 , Q Zhang1 ,D Nguyen1 ,W Chen2 , L Zhou3 , S Jiang1 , (1)Medical Artificial Intelligence and Automation Laboratory, Department of Radiation Oncology, UT Southwestern Medical Center, Dallas, Texas, 75235, (2)Department of Radiology, Nanfang Hospital, Southern Medical University, Guangzhou, China, 510515, (3)Institute of Medical Instrument, School of Biomedical Engineering, Southern Medical University, Guangzhou, China, 510515

Presentations

(Tuesday, 7/31/2018) 4:30 PM - 5:30 PM

Room: Exhibit Hall | Forum 2

Purpose: A deep learning model was developed for mass detection and segmentation to assist physicians for breast cancer screening and diagnosis with Digital Breast Tomosynthesis (DBT).

Methods: DBT is a newly developed three-dimensional (3D) imaging modality, which holds a great potential to improve the accuracy of mammography especially for dense breasts by reducing tissue overlap. While DBT increases the detection rate of mass-like lesions, much more manual detection and segmentation work need to be done by radiologists, which greatly reduces the efficiency. Therefore, a method that can accurately and automatically detect and segment mass is highly demanded in the clinical practice. In this study, a deep learning model based on U-Net was developed for solving this problem. A total of 964 cases with diagnosed mass-like lesions were retrospectively used for model training and testing. Patients with Infiltrating Ductal Carcinoma (IDC), Ductal Carcinoma in Situ (DCIS), Invasive Lobular Carcinoma (ILC), Adenofibroma, Cystic Hyperplasia, Cyst and Hyperplasia were included. The follow up time was about 24 months and malignant cases were confirmed by biopsy or surgical pathology. DBT images were contoured and reviewed by 3 radiologists, each with more than 5 years’ experience in breast radiology. The deep learning model was trained based on the DBT images with mass-like lesions as input, and radiologists’ contours as ground truth.

Results: In this work, detection rate, dice coefficient per case, and specificity were used as the evaluation criteria. The deep learning model shows a detection rate of 93.1%, and an average dice coefficient of 80.9%.

Conclusion: This study presented an end-to-end deep learning network for breast mass detection and segmentation from DBT images. Our method obtained very promising result on breast mass detection and segmentation and especially in dense breasts.

Funding Support, Disclosures, and Conflict of Interest: Supported by Science and Technology Planning Project of Guangdong Province, China(No.2016ZC0058) and Medical Scientific Research Foundation of Guangdong Province, China(No. A2017496).

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email