1. Quantitative Doppler Vascularity
Improves Computer-Based Sonographic
Diagnosis of Breast Cancer
Laith R Sultan MD, Santosh S Venkatesh PhD, Emily Conant
MD, Chandra Sehgal PhD
Ultrasound Research Lab., Department of Radiology, University of
Pennsylvania, Philadelphia PA
Contact e-mail: lsultan@mail.med.upenn.edu
3. Background and Objective
Background:
- Abnormal angiogenesis is commonly linked with malignant breast lesions.
- Current studies do not routinely utelize quantitative Doppler for lesion
characterization
Objective: To assess the role of quantitative Doppler ultrasound in enhancing
breast cancer diagnosis.
Approach: To use machine learning methods based on computerized training
and testing of image features for diagnosis
4. Grayscale (GS): 8 computerized GS features
describing lesion morphology and margin sharpness
-Angular variation
-Brightness difference
-Margin sharpness
-Axis ratio
-Depth to width ratio
-Radius variation
-Skelton norm
-Tortousity
-160 biopsy proven solid breast masses were studied for quanitative Grayscale
(GS) andColor Doppler (CD) features extracted from user defined margins
Materials and Methods
5. Materials and Methods
Color Doppler (CD):
• 3 vascularity features were analyzed
- vascular fraction area
- flow velocity index, &
- flow volume index.
• Vascular features were measured in 3
regions: center and rim of breast lesions
and surrounding area
• Lesion is described by user & computer
dervies outer and inner margins of equal areas
6. Materials and Methods
Data analysis
• GS features with patient age and CD features were used independently to train
the system using logistic regression classifier
• Cross validation was performed using the leave-one-out testing method.
• Dispersion of regression coefficients (learning ability) from the mean were
measured.
• Cases with highest dispersion represented weak learning (low confidence
diagnosis) and were pruned.
• Diagnostic performance were measured by the area
under ROC (Az). Sensitivity (Se) and specificity (Sp)
were measured at Youden index threshold for each ROC
curve.
7. Results
● Age of patients with benign lesions (45.9 ± 13.4 years) were significantly
different from malignant (57. 8 ± 12.1 years), (P < 10 e-7).
10. Results
Table 1: The diagnsotic performances for GS alone, GS plus
CD and selected cases.
Grayscale Grayscale +
Doppler
Selected cases
AUC (± SE) 0.85 ± 0.03
(0.79- 0.91)
0.89 ± 0.03
(0.83- 0.94)
0.96 ± 0.01
(0.92- 0.99)
Sensitivity 0.87 0.79 0.92
Specificity 0.69 0.89 0.95
11. Conclusion
• Adding Doppler vascularity features to sonographic morphologic features
markedly improves the diagnostic performance of machine learning
algorithms for breast cancer diagnosis.
• Our analysis identifed cases (12%; 18/160) that do not contribute to learning
( high dispersion), when pruned a very high diagnostic performance can be
achieved.