Abstract
Objective
Radiomic analysis of contrast enhanced mammographic (CEM) images is an emerging field. The aims of this study were to build classification models to distinguish benign and malignant lesions using a multi vendor dataset and compare segmentation techniques.
Methods
CEM images were acquired using Hologic and GE equipment. Textural features were extracted using MaZda analysis software. Lesions were segmented with freehand region of interest (ROI) and ellipsoid_ROI. Benign/Malignant classification models were built using extracted textural features. Subset analysis according to ROI and mammographic view was performed.
Results
269 enhancing mass lesions (238 patients) were included. Oversampling mitigated benign/malignant imbalance. Diagnostic accuracy of all models was high (>0.9). Segmentation with ellipsoid_ROI produced a more accurate model than with FH_ROI, accuracy:0.947 vs 0.914, AUC:0.974 vs 0.86, p < 0.05. Regarding mammographic view all models were highly accurate (0.947–0.955) with no difference in AUC (0.985–0.987). The CC-view model had the greatest specificity:0.962, the MLO-view and CC + MLO view models had higher sensitivity:0.954, p < 0.05.
Conclusions
Accurate radiomics models can be built using a real-life multivendor dataset segmentation with ellipsoid-ROI produces the highest level of accuracy. The marginal increase in accuracy using both mammographic views, may not justify the increased workload.
Advances in knowledge
Radiomic modelling can be successfully applied to a multivendor CEM dataset, ellipsoid_ROI is an accurate segmentation technique and it may be unnecessary to segment both CEM views. These results will help further developments aimed at producing a widely accessible radiomics model for clinical use.
Radiomic analysis of contrast enhanced mammographic (CEM) images is an emerging field. The aims of this study were to build classification models to distinguish benign and malignant lesions using a multi vendor dataset and compare segmentation techniques.
Methods
CEM images were acquired using Hologic and GE equipment. Textural features were extracted using MaZda analysis software. Lesions were segmented with freehand region of interest (ROI) and ellipsoid_ROI. Benign/Malignant classification models were built using extracted textural features. Subset analysis according to ROI and mammographic view was performed.
Results
269 enhancing mass lesions (238 patients) were included. Oversampling mitigated benign/malignant imbalance. Diagnostic accuracy of all models was high (>0.9). Segmentation with ellipsoid_ROI produced a more accurate model than with FH_ROI, accuracy:0.947 vs 0.914, AUC:0.974 vs 0.86, p < 0.05. Regarding mammographic view all models were highly accurate (0.947–0.955) with no difference in AUC (0.985–0.987). The CC-view model had the greatest specificity:0.962, the MLO-view and CC + MLO view models had higher sensitivity:0.954, p < 0.05.
Conclusions
Accurate radiomics models can be built using a real-life multivendor dataset segmentation with ellipsoid-ROI produces the highest level of accuracy. The marginal increase in accuracy using both mammographic views, may not justify the increased workload.
Advances in knowledge
Radiomic modelling can be successfully applied to a multivendor CEM dataset, ellipsoid_ROI is an accurate segmentation technique and it may be unnecessary to segment both CEM views. These results will help further developments aimed at producing a widely accessible radiomics model for clinical use.
Original language | English |
---|---|
Article number | 20220980 |
Number of pages | 8 |
Journal | The British Journal of Radiology |
Volume | 96 |
Issue number | 1145 |
Early online date | 6 Mar 2023 |
DOIs | |
Publication status | Published - 1 Apr 2023 |
Keywords
- Contrast-enhanced mammography
- Mammography
- Radiomics
- Breast cancer