ISRA May 2022

Application of Delayed-Contrast MRI for Improved Sensitivity in Breast Cancer Diagnosis

Dianne Daniels 1 David Last 1 Naomi Fridman 1 Nora Lahat 3 Renata Faermann Weidenfeld 1 Osnat Halshtok 1 David Samoocha 1 Yael Mardor 1,2 Miri Sklair-Levy 1,2
1Radiology, Sheba Medical Center, Israel
2Medicine, Tel Aviv University, Israel
3Pathology, Sheba Medical Center, Israel

Purpose: Contrast-enhanced MRI of the breast provides high sensitivity but variable specificity in detecting breast cancer, and may lead to excessive benign biopsies. Treatment response assessment maps (TRAMs) are calculated from delayed-contrast MRI and reflect delayed-contrast clearance/accumulation. In brain tumor patients, TRAMs provide high sensitivity/specificity (>90%) for differentiating tumor (contrast clearance, blue in TRAMs)/non-tumor (accumulation, red) tissues, currently in routine clinical use. Here, we studied their application for reducing benign biopsies in breast lesions.

Methods: 92 Women with 133 breast lesions suspected as tumors were scanned by standard (including DCE) and delayed-contrast MRI. Lesions were determined by biopsy(119)/follow-up(18) as malignant/benign. The performance of benign/malignant classifiers calculated from the TRAMs were studied using a supervised machine-learning (ML) algorithm and leave-one-out cross-validation (LOOCV).

Results: 100% of malignant lesions (78) appeared blue in the TRAMs while benign (55) lesions appeared mixed red/blue. The lesion features found to be most predictive were: blue portion, blue intensity and largest blue cluster. The TRAMs-based classifier combining these features resulted in: sensitivity=99%/specificity=60%/PPV=78%/NPV=97%/accuracy=83%, while standard DCE resulted in: sensitivity=100%/specificity=33%/PPV=68%/NPV=100%/accuracy=72%. Comparison of TRAMs to histology revealed different vascular patterns for blue (vessels with closed/compressed lumens) and red (open lumens) regions.

Conclusion: The TRAMs-based classifier may provide improved diagnosis of breast cancer as it resulted in nearly no false negatives and 60% of correctly identified benign tumors, thus allowing significant reduction of benign biopsies. Expanding to more advanced ML methods and a larger cohort is ongoing.

Limitations: Our cohort is naturally biased towards malignant lesions. Still, this does not affect the reliability determined by the negligible percentage of false negatives. False positive cases were mainly due to hemangiomas/fat necrosis/fibroadenomas with myxoid components.