Leithner et al. (2019) [9] |
143 |
First order histogram, co-occurrence matrix |
Luminal A, luminal B, TNBC, HER2-enriched |
Accuracy: luminal B vs. luminal A, 84.2%; luminal B vs. TNBC, 83.9%; luminal B vs. all others, 89%; HER2-enriched vs. all others, 81.3% |
Kim et al. (2020) [10] |
228 |
Radiomics score from 5 MRI features |
Disease free survival outcome of TNBC |
Combined clinicopathological and radiomics feature model showed highest AUC (0.844) in prediction of DFS |
Song et al. (2020) [11] |
92 |
GLCM features |
HER2-positive vs. −negative tumors determined by FISH test |
Among 3 machine learning methods, LRA, QDA, and SVM, AUC of SVM was the best (0.890) |
Mazurowski et al. (2014) [12] |
48 |
Semiautomatically extracted MRI features |
TCGA database contains full genomic sequencing |
Higher ratio of lesion enhancement rate to background parenchymal enhancement was more likely to be luminal B subtype |
Grimm et al. (2015) [13] |
275 |
56 Features (morphologic, texture, and dynamic features) |
luminal A, luminal B, HER2, basal |
Luminal A and luminal B molecular subtype were associated with semiautomatically extracted imaging features |
Bae et al. (2015) [14] |
280 |
Tumor roundness score |
ER, PR, Ki-67 |
ER score and Ki-67 index were independent factors determining tumor roundness. TNBC showed the highest mean roundness scores compared with the other subtypes |
Sutton et al. (2015) [15] |
95 |
Morphology, histogram features, GLCM features |
Oncotype DX recurrence score of Luminal A tumor |
An increased kurtosis was found to be a statistically significant factor correlating with Oncotype DX recurrence score |
Waugh et al. (2016) [16] |
221 |
Texture features |
Hormone receptor–positive and −negative cancers demonstrated significantly different entropy features |
Textural differences on contrast-enhanced MR images might reflect underlying lesion subtypes |
Li et al. (2016) [17] |
91 |
Texture features |
Luminal A, luminal B, HER2-enriched, and basal-like subtype |
Enhancement texture (entropy) and molecular subtypes were related. AUC of ER+ vs. ER−: 0.89, PR+ vs. PR−: 0.69, HER2+ vs. HER2−: 0.65, TNBC vs. others: 0.67 |
Agner et al. (2014) [18] |
76 |
Morphology, kinetic intensity, histogram features, GLCM features |
TNBC, HER2, ER-positive tumors |
AUC of texture features (more heterogeneity) for TNBC vs. other subtypes: 0.73 to 0.74 |
Chamming’s et al. (2018) [19] |
85 |
Fine, medium, and coarse texture for mean, standard deviation, mean proportion of positive pixels, entropy, skewness, and kurtosis |
pCR of TNBC |
kurtosis appears to be associated with pCR to neoadjuvant chemotherapy in non–TNBC and may be a promising biomarker for the identification of TNBC |
Braman et al. (2017) [20] |
117 |
Intratumoral and peritumoral texture features |
Prediction of pCR |
Combined intratumoral and peritumoral radiomics for prediction of pCR yielded AUC 0.83 for HR+ HER2−. Non-pCR was characterized by elevated peritumoral heterogeneity during initial phase. For TNBC/HER+ tumors were best characterized by a peritumoral speckled enhancement. |