Click here to


Are you sure ?

Yes, do it No, cancel

Breast Ultrasound Computer-Aided Diagnosis Using Triplet Path Networks with Fisher Discriminant Analysis

E Zhang1*, Z Yang2 , S Seiler3 , S Yu4 , M Chen5 , W Lu6 , X Gu7 , (1) UTSouthwestern Medical Center, Dallas, TX, (2) UTSouthwestern Medical Center, Dallas, TX, (3) UTSouthwestern Medical Center, Dallas, TX, (4) Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, Guangdong, (5) UT Southwestern Medical Center, Dallas, TX, (6) UT Southwestern Medical Center, Dallas, TX, (7) UT Southwestern Medical Center, Dallas, TX


(Wednesday, 7/17/2019) 1:45 PM - 3:45 PM

Room: 303

Purpose: To develop a novel BI-RADS features oriented triplet path networks, namely BIRADS-TPN, for accurate breast ultrasound (US) computer-aided diagnosis (CAD).

Methods: The proposed BIRADS-TPN is designed to incorporate clinical-approved breast lesion characteristics (BI-RADS features) and a triplet path networks with Fisher discriminant analysis (TPN) to achieve accurate diagnosis on US images with small training dataset. Original breast US images are converted to BIRADS-oriented feature maps (BFMs) using a distance-transformation coupled Gaussian filter. Then the converted BFMs are used as the input of TPN algorithm, which is characterized as Fisher discriminant analysis guided two unsupervised stacked convolutional Auto-Encoder (SCAE) networks for benign and malignant image reconstruction independently. The developed BIRADS-TPN algorithm is trained with an alternative learning strategy by balancing reconstruction error and Fisher loss to ensure the representation and discriminancy of features from the shared sub-network. Original images based (ORI-) or BIRADS based (BIRADS-) TPN are compared to conventional neural network-based lesion classification using features from unconstrained SCAE. Networks are trained and tested on 295 patients’ ultrasound images (205 benign and 90 malignant) with 80% for training and the remaining for testing.

Results: The performance of trained networks are evaluated with metrics, including accuracy (ACC), area under receiver operating characteristic curve (AUC), sensitivity (SEN), specificity (SPE). Evaluation results show BIRADS-TPN achieved the best performance among comparison with ACC and AUC ~83 and 76%, respectively.

Conclusion: BIRADS-TPN achieved high accuracy on the testing dataset and ranked the best performance among four comparing networks, which indicated that BIRADS-TPN could a promising model for effective US breast lesion CAD using small datasets.

Funding Support, Disclosures, and Conflict of Interest: This work was supported by the Seed Grant from Department of Radiation Oncology at University of Texas Southwestern Medical Center.


Breast, Ultrasonics, CAD


IM- Ultrasound : CAD

Contact Email