LAUSR

Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is the most sensitive imaging modality in the detection of breast cancer, with reported sensitivity ranging from 89–100%. Indications for DCE-MRI include breast cancer screening in women at increased risk and assessment of extent of disease in women with known breast cancer. However, a limitation of DCE-MRI is its moderate specificity (ranging from 35–64%), leading to false positives, benign biopsies, and possible overtreatment. Therefore, many investigators have sought to improve lesion classification to differentiate benign from malignant lesions, with promising increases in accuracy with radiomics techniques and even better accuracies with deep-learning techniques. In the current issue of JMRI, Zhou et al compare characterization of benign and malignant lesions on DCE-MRI using conventional methods (analyzing 3D volume and washin and washout kinetics), a radiomics model, and deeplearning models. Their findings confirm those of previous authors showing that deep-learning models outperform the radiomics and conventional models. Further, Zhou et al compared five different deep-learning models, which differ by the size of the input box containing the lesion as well as varying amounts of perilesional tissue. The authors found that, although the tumor-only model had high sensitivity (100%), the specificity rose significantly (from 61–79%) when a small amount of perilesional tissue was included in the input. This confirms the importance of the perilesional tissue in the diagnosis of breast cancer. Molecular biology studies have demonstrated that the tumor microenvironment can profoundly influence tumor behavior, metastasis, and patient outcomes. The interplay of breast cancers and their environment are manifold. Examples include collagen deposition increasing around growing tumors and facilitating invasion, and peritumoral lymphatic vessel density correlating with lymph node metastasis. The importance of the peritumoral space has recently been described in the radiology literature. Cheon et al demonstrated that peritumoral edema is itself a prognostic feature. In an important study, Braman et al demonstrated that radiomics features from peritumoral regions discriminated HER2+ breast cancer subtypes, were associated with response to HER2-targeted therapy, and were associated with the density of tumorinfiltrating lymphocytes. Interestingly, Zhou et al found that including larger amounts of perilesional tissue with larger bounding boxes decreased the accuracy of the model. However, there is evidence that even apparently normal fibroglandular tissue remote from a known lesion can be predictive. Kim et al incorporated features from tumor and nontumor parenchyma and found that the K in the nontumor tissue was significantly different between malignant and benign groups, and was as predictive of malignancy as the K within the lesion itself. This suggests that breast cancer may have an effect on the entire breast, which is an area deserving future study. Zhou et al included only mass lesions, which was done in order to ensure a sharp lesion boundary to ensure sufficient inclusion of perilesional tissue. It is perhaps because of the inclusion of only masses that the majority of cancers were invasive ductal carcinomas (82% in the training set and 70% in the test set), with a minority of ductal carcinoma in situ (DCIS) and other invasive cancers. There is, however, a clear need for improved characterization of nonmass lesions, as invasive lobular carcinoma and DCIS can mimic benign entities such as fibrocystic change and background enhancement. Therefore, future studies should include a wide range of benign and malignant lesions, as well as mass and nonmass lesions, to simulate real-world applications. Incorporation of multiparametric data into deeplearning models is another area for future research. Zhou et al note that they used DCE-MRI inputs only, and the addition of T2-weighted images could have provided helpful diagnostic information. Diffusion-weighted imaging has also been found to significantly improve lesion characterization. The addition of diffusion-weighted imaging to multiparametric deeplearning models has already been shown to improve accuracy of lesion characterization.