AbstractIntroduction: Contrast-enhanced digital breast tomosynthesis (CE-DBT) is a novel imaging technique, combining contrast enhanced mammography (CESM) with tomosynthesis, which may offer an alternative to MRI for monitoring response to neoadjuvant chemotherapy (NACT). Textural analysis (TA) may increase CESM accuracy, generating quantitative functional data.
Methods: CONDOR was a prospective feasibility study in which women undergoing NACT were imaged with CE-DBT and MRI. Patient experience was investigated using questionnaires. The separate and combined components of CE-DBT were assessed; ie. Low energy (LE) mammogram, tomosynthesis (DBT), CESM contrast-enhancement (CESM(CE)), combined CESM contrast and microcalcifications seen on LE (CESM(CE+calc) and the fully combined CE-DBT. Imaging techniques were directly compared for accuracy identifying multifocality, complete pathological response (pCR), residual whole tumour size (WTS) and invasive tumour size (ITS). Additional patients’ CESM images were acquired using either Hologic or GE equipment. Lesions were segmented with freehand-ROI and ellipsoid-ROI on craniocaudal (CC) and mediolateral oblique (MLO) views and textural features were extracted. Machine learning was used to build models for the classification of lesions according to malignancy, tumour grade, receptor status and pCR-status. CESM-washout features were calculated by comparing enhancement on initial and delayed-MLO views. These were correlated to MRI-curves.
Results: Eighteen patients with 24 cancers were enrolled. Response prediction accuracy for pCR was 81.25% (95%CI: 54.35-95.95) for CESM(CE) vs 62.5% (95%CI: 35.43-84.80) for MRI; ITS concordance coefficient was 0.70 for CESM(CE) vs 0.66 for MRI; WTS concordance coefficient was 0.69 for CESM(CE+calc) vs 0.87 for MRI. Accuracy for identifying multifocality was 94.6% (95%CI:81.8-99.3) for CESM(CE) vs 89.2%(95%CI:74.6-97.0) for MRI. No benefit from the addition of DBT was identified for any measure. On 77% of occasions patients preferred CE-DBT, with significantly better overall experience (p=0.008).
All TA models for classifying malignant lesions were highly accurate. The Ellipsoid_ROI model performed better than FH_ROI; sensitivity 0.998 vs 0.953, specificity 0.916 vs 0.891, p<0.05. Both-view model produced most consistently good results, followed by CC-view. There was no difference in AUC; 0.987 vs 0.988 vs 0.985 for both-view vs CC-view vs MLO-view. Promising results were seen for prediction of ER-status. Model accuracy was poorer for prediction of tumour grade, PR-status, HER-status and pCR status, potentially due to small subset numbers. CESM-washout features generated from ellipsoid_ROI were significantly different between the MRI curve types.
Conclusion: I have demonstrated that CESM has similar accuracy to MRI for identifying multifocality and predicting response to NACT, and is preferred by patients. The combination of CESM and low-energy mammogram – CESM(CE+calc) – increases the sensitivity for residual in situ disease. No additional benefit was afforded by combining CESM with tomosynthesis.
I have demonstrated the first multivendor CESM classification model using textural analysis data to identify malignant lesions. Promising results were seen for prediction of ER-status. With larger datasets this may translate to more accurate pCR classification models. Successful quantification of CESM-washout was achieved with significant differences between MRI curve types. I recommend further work utilises segmentation Ellipsoid_ROIs on single-view CESM images.
|Date of Award||2022|
|Supervisor||Andrew Evans (Supervisor), Graeme Houston (Supervisor) & Sarah Vinnicombe (Supervisor)|