NAC is a central element of contemporary breast cancer management, however there is a wide spectrum of response to NAC between patients which can vary based on host, tumour and treatment factors. Patients with locally advanced breast cancer and selected patients with early stage breast cancer are offered NAC with the aim of down-staging the tumour size or gaining valuable information relating to in-vivo tumour response. (30). Appropriate patient selection for NAC is vital and can be informed by gene expression biomarkers and indexes of proliferation such as ki67. Recent evidence suggests that radiomic features from pre-NAC MRI imaging can be utilised to non-invasively predict response to NAC (31–35) and potentially contribute to the existing patient-selection paradigm.
This study has identified 4 radiomic features from pre-NAC MRI to stratify a cohort of 74 patients with invasive breast cancer into poor and excellent response groups. A machine learning approach was utilised to select these pertinent features and to build a model to predict response to NAC. The addition of estrogen receptor status improved the overall performance of the model, with an AUC of 0.811, identifying poor responders with 90% sensitivity and 70% specificity.
Conventional molecular subtypes help inform likelihood of response to NAC. A study of 838 patients demonstrated significantly different rates of pCR between Luminal A, Luminal B, HER2 over-expressing and Triple-Negative subtypes (6%, 16%, 37% and 38% respectively) (36). The results of this study compare favourably to similar studies investigating the role of biomarkers in predicting response to NAC. Oncotype DX has been validated in adjuvant therapy, however its’ role in NAC is less clear; a 2019 study of 989 breast cancer patients found that a high Oncotype recurrence score (>30) was significantly associated with pCR (Odds ratio 4.87) (37). OncoMasTR is another multi-gene prognostic signature in development specifically to predict NAC response and incorporates 3 master transcription regulator genes as well as tumour size and nodal status (38). A 2020 study of 813 breast cancer patients showed that OncoMasTR score was significantly associated with pCR (OR 1.68) (39).
Recent evidence supports the addition of radiomic features as potential predictors of NAC response. A 2019 study investigating histopathological residual cancer burden in 38 breast cancer patients utilised 23 pharmacokinetic features obtained from DCE-MRI in addition to conventional pathological factors to classify response to NAC with an AUC of 0.92 (34). A 2020 study of 222 breast cancers utilised a model composed of 12 MRI-derived radiomics features in addition to molecular subtype to identify pCR, producing an AUC of 0.8 using a random forest machine learning approach (31). In cervical cancer, a 2019 study of 275 patients demonstrated an AUC of 0.999 in predicting response to NAC (40).
The radiomic features identified and tested may further describe the intrinsic tumour environment and the degree of intra-tumour heterogeneity which may impact NAC response. Kurtosis can be used as a measure to assess deviation from the normal distribution of pixel values. Invasiveness may be explained by the degree of pixel-kurtosis in breast cancer (41) and has been shown to be associated with response to chemotherapy in pancreatic cancer (42). Texture based features evaluating the relationship between pixels are produced by using spatial grey-level dependant matrices. NGLDM (neighbourhood grey-level different matrix) describes the difference in grey-levels between 1 voxel and it’s 26 neighbours in 3 dimensions. NGLDM contrast corresponds to intensity difference between neighbouring regions, and this is the first report of this radiomic feature in association with cancer prognosis or response to chemotherapy.
Grey-level zone length matrix (GLZLM) provide information on the size of homogenous zones for each grey level in 3 dimensions. GLZLM_SZE (short zone emphasis)is a measure of the distribution of the short homogenous zones in an image, while GLZLM_ZP (zone percentage) measures the homogeneity of the homogenous zones. Indices derived from the GLZLM, in addition with kurtosis, were significantly associated with overall survival in a study investigating radiomics features in gastric B-cell lymphoma (43). A 2020 study that assessed PET scan radiomic features of patients with pancreatic cancer demonstrated that GLZLM non-uniformity was significantly associated with one-year survival and could stratify patients into survival categories (44).
Radiomic features alone show great promise in the stratifying response to NAC, and models incorporating a combination of radiomics and molecular feature are superior (45, 46). It is conceivable that radiomic features could be a component of future multi-omic panels including genomic and metabolomic markers to aid in the management of breast cancer (47, 48). We added ER status to moderately improve the overall accuracy of the model to predict response to NAC and produce a greater accuracy in classifying patients into poor and excellent response groups than conventional molecular subtype alone. However, response to NAC can vary significantly even within subtype and is thought to be as result of intra-tumour heterogeneity (49). Genomic heterogeneity has been shown to impact treatment response and drive resistance to targeted therapies in cancer (50, 51). Image-based assessment of tumour heterogeneity, incorporating quantitative descriptors of grey-level relationships mentioned above, could potentially reveal aggressive tumour sub-regions for determining prognosis and treatment (52, 53) and be incorporated into the multi-modal decision process of selecting patients for NAC.
This study has a number of limitations. Firstly, it is a single centre retrospective study. While we were able to establish a discrete number of radiomic features to predict response to NAC, a larger sample size is needed to validate the radiomic model. Because radiomics is itself a developing field, there is a paucity of large cohort, prospective studies assessing the clinical utility of radiomic models. Establishing a robust, reproducible radiomics pipeline as is demonstrated in this study, is vital to integrate radiomic biomarkers into clinical practice in the near future (54).
In our study, tumour segmentation was carried out manually by a single researcher, under the supervision of a Consultant Radiologist. Manual segmentation is at present the most reliable method of establishing a region of interest (ROI) for analysis by radiomics software (15). However, this method can be subject to inter-observer variability. Automatic segmentation by artificial intelligence shows great promise in solving this issue however is some way from being optimised (55).
Pre-processing of images by resampling has been shown to reduce more repeatable and less sensitive to change results (56, 57). Here, we used a fixed bin width of 64 and carried out voxel resampling to isotropic voxels of size 2 x 2 x 2 mm3 by 3-dimensional Lagrangian interpolation, as described in previous studies assessing breast MRI (58, 59). Optimal and standardized pre-processing must be established to ensure reproducibility across radiomics studies (60).
In terms of feature extraction software, LIFEx was utilised that produces radiomics features compliant with the International Biomarker Standardization Initiative (IBSI) (61). This 2020 initiative describes 169 standardized radiomics features that are reproducible across a number of software platforms and can potentially be investigated as clinical biomarkers. As radiomics evolves in the coming years it is likely that the number of software programmes available will increase and it is imperative that rigorous assessment of features continues to ensure reproducibility and reliability of studies.
In conclusion, this study identifies radiomic features that could potentially contribute to the management of patients receiving NAC for breast cancer.