Three-Compartment Breast Lesion Detection (3CB)

The Shepherd Research Lab has developed a novel breast-imaging technique that analyzes lesion composition using three compartmental measurements: protein, lipid, and water (3CB). The study investigates whether lipid-protein-water signatures of mammographically suspicious breast lesions can improve cancer diagnosis and reduce unnecessary biopsies.

The first five years of research (March 2013 – June 2017) are complete. Current funding focuses on developing and evaluating novel mammographic biomarkers combined with quantitative image analysis (QIA) and radiomics methods — collectively termed q3CB.

Aim 1 — Sensitivity & specificity of 3CB signatures

Recruitment adjusted from 600 to 498 FFDM patients due to a 17% budget reduction. As of the last update, 425 women enrolled (215 UCSF / 210 Moffitt) with biopsy-confirmed subtypes: 61 invasive ductal carcinoma (IDC), 40 ductal carcinoma in situ (DCIS), 66 fibroadenomas, 324 benign findings.

A calibration phantom with 51 combinations of water, lipid (wax), and protein (Delrin) was created for standardizing 3CB across FFDM and DBT systems.

Early analysis of 45 lesions demonstrated 3CB features distinguishing between lesion types:

• IDC from DCIS by lipid skewness — AUC = 0.71
• Fibroadenomas by water texture relative to background — AUC = 0.75
• Benign lesions by peripheral water content — AUC = 0.71

Combined 3CB signature achieved AUC = 0.72 in cross-validation for invasive-cancer detection.

Aim 2 — Comparison with CAD/QIA methods

Merged QIA/radiomics signatures yielded AUC = 0.81 in distinguishing lesions requiring biopsy from those not requiring it. Key QIA features identified:

• Invasive cancer: spiculation
• DCIS: circularity
• Fibroadenoma: radial gradient index
• Other benign findings: texture heterogeneity

Deep-learning approaches for microcalcification classification showed promising results. In a dataset of 99 biopsy-proven lesions, the deep-learning method “could have avoided 21 biopsies of the 80 benign lesions…versus only 8 avoidable biopsies based on radiologists (p < .001).”

Aim 3 — Combined 3CB and QIA performance

The combination of 3CB and QIA (q3CB) significantly improved classification:

Lesion type	3CB alone	QIA alone	Combined
Overall	0.71	0.81	0.86
Masses	—	0.83	0.89
Microcalcifications	—	0.84	0.91
Asymmetry / arch. distortion	—	0.61	0.87 (p = 0.006)

3CB compositional information and QIA features provide complementary diagnostic information with little correlation between them.

Future directions

Extension to 3D tomosynthesis imaging and reader studies to validate whether 3CB knowledge influences radiologist decision-making and reduces unnecessary biopsies.