TL;DR
This study explores the potential of radiomic analysis of contrast-enhanced mammography (CEM) for classifying breast lesions. They found that using high-energy (HE) images alone, or combining all three CEM image types (HE, low-energy (LE), and dual-energy subtraction (DES)), yielded the best classification performance, suggesting that even clinically uninterpreted HE images hold valuable diagnostic information.
What Problem Does This Paper Address?
Accurate and early diagnosis of breast cancer is critical. While mammography is a standard tool, its sensitivity can be limited, especially in women with dense breasts. Contrast-enhanced mammography (CEM) offers improved lesion visualization by highlighting angiogenic characteristics, producing three types of images: high-energy (HE), low-energy (LE), and dual-energy subtraction (DES). However, the HE images are not typically used for direct clinical interpretation. Radiomics, a field that extracts quantitative features from medical images, has shown promise in various cancer diagnostics. This paper addresses the gap in understanding how radiomic analysis of CEM, particularly incorporating the underutilized HE images, can improve the classification of benign versus malignant breast lesions.
Methods
The researchers employed a retrospective, double-center study design. They extracted radiomic features from HE, LE, and DES CEM images. To build classification models, they used the Maximum Relevance Minimum Redundancy (mRMR) method for feature selection, followed by Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression to identify the most predictive features and construct the models. The models' performance was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC). To ensure robustness, they also performed 100 rounds of Leave-Group-Out Cross-Validation (LGOCV).
Datasets
The study utilized CEM images from 192 women with 226 breast lesions, collected between November 2013 and February 2020 from two institutions. The dataset was split into training (70%) and testing (30%) sets. The lesions were categorized as either benign (101) or malignant (125). No new benchmarks were introduced; the study focused on analyzing existing CEM image data.
Key Insights and Ideas
- HE Images Hold Diagnostic Power: Clinically uninterpreted HE images, when analyzed radiomically, performed exceptionally well in classifying breast lesions, even outperforming LE and DES images individually. This suggests that radiomics can unlock valuable diagnostic information from images that are not typically used for direct human interpretation.
- Combining Image Modalities is Best: The highest classification performance was achieved when radiomic features from all three CEM image types (HE, LE, and DES) were combined. This highlights the complementary nature of the information captured by each image type.
- Radiomics as a Complementary Tool: The study demonstrates that radiomic analysis of CEM can provide a valuable, non-invasive method to aid in the differentiation of benign and malignant breast lesions, potentially reducing the need for unnecessary invasive biopsies.
- Robustness of Findings: The use of LGOCV across 100 rounds provided stable and reliable performance metrics, reinforcing the validity of the developed models.
Key Results
The best classification models achieved high AUCs:
- HE images alone: AUC of 0.931 (testing set)
- LE images alone: AUC of 0.897 (testing set)
- DES images alone: AUC of 0.882 (testing set)
- Combination of HE, LE, and DES images: AUC of 0.960 (testing set)
The combined model (HE + LE + DES) consistently showed the highest performance across various metrics, including AUC, accuracy, sensitivity, and specificity. Even the HE-only model demonstrated strong performance, outperforming LE and DES individually.
Limitations and Future Work
The authors acknowledge several limitations:
- Data Mixing: Data from two research centers were mixed for training and testing, rather than using a completely independent external validation set.
- Retrospective Nature: The study was retrospective, and some HE images were missing, leading to a smaller sample size than ideal.
- Manual Segmentation: Lesion segmentation was performed manually, which can introduce variability and affect repeatability.
- Sensitivity vs. Human Readers: The sensitivities of the radiomic models were lower than those reported for human readers, suggesting room for improvement.
- Dataset Bias: The dataset was enriched for malignant lesions, which might lead to an overestimation of model performance.
Future work could involve prospective studies with larger, more diverse datasets, exploring automated segmentation techniques, and further investigating the specific contributions of HE images to improve specificity.
Summary updated