Automated Segmentation of Suspicious Breast Masses from Ultrasound - - PowerPoint PPT Presentation

Viksit Kumar, Jeremy Webb, Adriana Gregory, Mostafa Fatemi, Azra Alizad

Automated Segmentation of Suspicious Breast Masses from Ultrasound Images

GTC 2018, San Jose, USA

2

SIGNIFICANCE

Breast cancer is most common and leading cause of death in American women* Segmentation and classification of suspicious breast masses can avoid unnecessary biopsies Conventional segmentation algorithm requires an initial seed Sørensen–Dice performance of 0.82-0.86, a measure of similarity of segmentation

Core Needle biopsy for suspicious breast masses

*Siegel RL, Miller KD, Jemal A. Cancer statistics, 2016. CA: a cancer journal for clinicians. 2015 *Shulman LN, Willett W, Sievers A, Knaul FM. Breast Cancer in Developing Countries: Opportunities for Improved Survival. Journal of Oncology. 2010;2010. doi: 10.1155/2010/595167

GOALS

• Develop an automated real-time detection and segmentation of

suspicious breast masses

• Currently ultrasonographers use the morphological and textural

features to identify suspicious masses based on suggestions from American College of Radiology

• The suspicious masses are then scored based on the Breast Imaging

Reporting and Data System (BI-RADS) scale and is the basis for recommending core needle biopsy

• Automated detection process can aid in rapid localization of suspicious

masses

• Automated segmentation and classification process can assist in

assigning BI-RADS score and recommending biopsy

4

ULTRASOUND IMAGING

Transducer Raw channel RF data Beam formed RF data In-phase quadrature data B-mode image Post processed B-mode image

MATERIALS AND METHODS

• 258 patients, 433 images from multiple cross-sections, 148 cases
• f BI-RADS 4
• 124 malignant and 134 benign cases
• Equal number of images from General Electric LOGIQ E9 and

Philips IU22

• Algorithm: Multi U-net with majority voting
• Loss function: negative dice coefficient
• Testing in real-time on Titan xp (13.83 ms/image)

6

ARCHITECTURE

Ronneberger O, Fischer P, Brox T, editors. U-net: Convolutional networks for biomedical image segmentation. International Conference on Medical Image Computing and Computer-Assisted Intervention; 2015: Springer

SEGMENTATION RESULTS-CONTD.

B-mode image of a benign-cellular fibro epithelial mass. The mass has typical smooth boundaries of a benign mass and is oval in shape

Red outline = Expert segmentation Blue outline = Predicted segmentation

2(b)

Dice coefficient = 0.94

SEGMENTATION RESULTS-CONTD.

B-mode image for benign fat necrosis with dystrophic calcifications. Notice the posterior acoustic shadowing beneath the benign mass

Red outline = Expert segmentation Blue outline = Predicted segmentation

3(a) 3(b)

Dice coefficient = 0.88

SEGMENTATION RESULTS-CONTD.

The biopsy result from the suspicious mass was malignant invasive/infiltrating ductal carcinoma, grade III. The irregular boundaries of the mass are typical of malignant masses which are usually challenging cases for segmentation algorithms

Red outline = Expert segmentation Blue outline = Predicted segmentation

4(a) 4(b)

Dice coefficient = 0.88

RESULTS: CINE CLIP 1

Benign fibroadenoma

• CONTD. CINE CLIP 2

Invasive Ductal carcinoma G.II/III

• Mean Sørensen–Dice coefficient of 0.82
• Real-time application with no need for initial seed
• Can be used for segmentation and detection of suspicious breast masses
• Segmentation masks can be used for classification of suspicious masses
• Reduce localization time and aid sonographer in categorizing the

suspicious mass

• Applications in point of care, mobile health monitoring, assisting

sonographers

• Provides the expertise of an expert sonographer to sonographers in

training

CONCLUSION

13

• We gratefully acknowledge the support of NVIDIA Corporation

with the donation of the Titan X Pascal GPU used for this research

• We gratefully acknowledge the support of Amazon web services

for the donation of credits used for this research

• This work was supported by National Institute of health grants

R01CA148994 from the National Cancer Institute

ACKNOWLEDGEMENT