Fig 1.
Representative B-mode and QUS spectral parametric images of ASD, AAC, MBF, SS, and SI from A benign (left three columns) and B malignant (right three columns) breast lesions.
The colour-bar range is 160 μm for ASD, 70 dB/cm3 for AAC, 20 dB for MBF, 10 dB/MHz for SS, and 70 dB for SI. The scale bar represents 1cm. The benign breast lesions were diagnosed as fibroadenomas, and complicated a cyst, respectively. The malignant lesions were diagnosed as invasive ductal carcinomas (IDC), invasive mammary carcinoma, and invasive lobular carcinoma (ILC). Using these parametric images, mean-value, textural, and image quality features were determined as imaging biomarkers for the characterization of breast lesions.
Fig 2.
Representative box and scatter plots of features that demonstrate statistically significant difference (p-values < 0.05) between benign (‘B’) and malignant (‘M’) lesion groups.
The first row shows core and margin mean-value parameters. The second row depicts representative core and margin GLCM features that showed discriminative power. The third row shows representative core and margin GRLM features that discriminate the two lesion groups. The last row depicts representative GLSZM features that provided the most discriminative power. There is a total of 160 features from tumour core and 5-mm margin, including 10 image quality features, available for feature selection. Among these features, 4 mean-values, 125 textural, and 1 image quality features demonstrate statistically significant difference between the two lesions. Statistically significant (p < 0.05), highly significant (p < 0.01), and extremely significant (p < 0.001) are shown with (*), (**), and (***), respectively.
Table 1.
Optimum feature set for classification using both core and margin information utilizing GLCM, GRLM, and GLSZM texture methods and SVM-RBF classification algorithm.
A maximum of 10 features was selected for classification. Model performance was evaluated using LOOCV method. Features were selected using forward SFS based on F1-score metric. Textural features, for example Core-MBF-CON: GLCM contrast parameter of MBF parametric image from core ROI and Margin-MBF-SALGE: GLSZM small area low gray level emphasis parameter of MBF parametric image from margin ROI, were the dominant features that contributed to hybrid biomarkers that best separated the two lesion types.
Table 2.
A: Core classification results of GLCM methodology using LOOCV. B: Core classification results of GLCM methodology using hold-out validation.
Table 3.
A: Margin classification results of GLCM methodology using LOOCV. B: Margin classification results of GLCM methodology using hold-out validation.
Table 4.
A: Core and margin classification results of GLCM methodology using LOOCV. B: Core and margin classification results of GLCM methodology using hold-out validation.
Table 5.
A: Core classification results of GRLM methodology using LOOCV. B: Core classification results of GRLM methodology using hold-out validation.
Table 6.
A: Margin classification results of GRLM methodology using LOOCV. B: Margin classification results of GRLM methodology using hold-out validation.
Table 7.
A: Core and margin classification results of GRLM methodology using LOOCV. B: Core and margin classification results of GRLM methodology using hold-out validation.
Table 8.
A: Core classification results of GLSZM methodology using LOOCV. B: Core classification results of GLSZM methodology using hold-out validation.
Table 9.
A: Margin classification results of GLSZM methodology using LOOCV. B: Margin classification results of GLSZM methodology using hold-out validation.
Table 10.
A: Core and margin classification results of GLSZM methodology using LOOCV. B: Core and margin classification results of GLSZM methodology using hold-out validation.