(D-130) Using Radiomics Features Extracted from Multi-Parametric MRI to Help Predict Breast Cancer Response to Neoadjuvant Therapy

Introduction::

Predicting the response of breast cancer to therapy in an individual patient is difficult due to the uncertainty of how any given patient may respond to a particular course of treatment.  Neoadjuvant Therapy (NAT) for breast cancer includes chemotherapy, hormone therapy, and targeted therapy.  The main goal of NAT is to improve patient outcomes by reducing the size of the cancerous lesion prior to surgery.  Post-NAT patients can be separated into two main categories based on their individualized tumor response; pathological complete response (pCR), where residual cancerous tissue in the breast or axilla is absent; and non-pathological complete response (non-pCR), where there remains residual cancerous tissue. pCR is prognostic for improved long-term survival outcomes and is therefore an important metric to predict in order to guide escalation/de-escalation treatment strategies.

In this study, we extract radiomics features from multi-parametric MR imaging data of patients with breast cancer undergoing NAT at three different time points, pre-NAT (T0), 3 weeks following the initiation of NAT (T1), and 12 weeks following the initiation of NAT (T2).  The data is then interpreted to evaluate lesions and predict the patient’s response to NAT as either non-pCR or pCR.  The goal of this study is to establish the utility of radiomics features, with regard to predicting the outcome of NAT in a population of breast cancer patients.

Materials and Methods::

The images used in this study were from the ACRIN-6698/I-SPY-II TRIAL, with a patient sample size of 266.  The focus of this study was on the analysis of apparent diffusion coefficient (ADC) calculated from diffusion-weighted MRI and dynamic contrast enhanced (DCE)-MRI.  For each patient, at each specific time point, a region of interest (ROI) was pre-determined by the ACRIN-6698 trial and a segmentation was created for each respective lesion.  Each volume image was then bias field corrected and intensity normalized via histogram matching to a common reference, using the Insight Toolkit (ITK) Python package [1].  After the images were normalized, they were aligned with their corresponding segmentations so that radiomics data could be extracted.  In this study, Py Radiomics [2] was used to calculate the radiomics features of each image/segmentation pairing, where 107 radiomics features were then extracted for statistical analysis.

Results, Conclusions, and Discussions:: Results: Each radiomics feature was then subjected to a paired t-test from T0 to T1, and from T0 to T2.  From the 107 features extracted, 39 features met the criteria of being statistically significant for both  pCR and non-pCR (p-value < = 0.05) and being within plus or minus one standard deviation.  Of those 39 features, four met the above criteria for all the image types that were tested.  The four identified features are as follows: Gray Level Co-occurrence Matrix (GLCM) inverse difference normalized; GLCM inverse difference moment normalized GLCM correlation; and the original shape flatness. Where the above feature had the following p values amongst non-pCR patients: p = 2.23 E-12, p = 8.07E-09, p = 6.15E-08, p = 1.38E-07.  Those features then had the following p values amongst pCR patients: p = 2.27E-09, p = 9.29E-08, p = 4.57E-05, p =3.98E-07, respectively.
Discussion: Of the four features that were identified in this study, three of them fall under the GLCM texture feature category.  GLCM is a mathematical representation of the ROI based on the intensity values of the gray levels within the ROI.  Specifically, the GLCM correlation feature displays the linear dependency of gray level values to their respective voxels in the GLCM.  Where a high correlation between neighboring voxels is indicative of a more organized and predictable lesion texture, and a lower correlation is indicative of a more complex and irregular lesion texture.
Conclusions: We find that radiomics features extracted from multi-parametric MRI data from patients with breast cancer undergoing NAT provides statistically significant metrics for predicting ultimate pCR/non-PCR status.  Radiomics feature data shows significant promise to be used in conjunction with other existing breast cancer response prediction tools, such as mechanistic biophysical modeling, to create a more comprehensive suite of tools for the prediction of NAT response in breast cancer patients. 




Acknowledgements (Optional): :

This project was supported in part by the NIH/NLM (R25 LM014214) in the Department of Biomedical Engineering and Center for Biomedical Informatics at Wake Forest University School of Medicine, NSF REU Site (Award #1950281) in the Department of Biomedical Engineering at Wake Forest University School of Medicine, NIH-NCI K25CA204599, NIH-NCI P30CA012197, Wake Forest Baptist Medical Center Comprehensive Cancer Center Signaling and Biotechnology program pilot grant, and Wake Forest University School of Medicine Faculty Start-up Funds. 

References (Optional): :

[1] B. C. Lowekamp, D. T. Chen, L. Ibáñez, D. Blezek, "The Design of SimpleITK", Front. Neuroinform., 7:45. https://doi.org/10.3389/fninf.2013.00045, 2013.

[2] van Griethuysen, J. J. M., Fedorov, A., Parmar, C., Hosny, A., Aucoin, N., Narayan, V., Beets-Tan, R. G. H., Fillion-Robin, J. C., Pieper, S., Aerts, H. J. W. L. (2017). Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Research, 77(21), e104–e107. https://doi.org/10.1158/0008-5472.CAN-17-0339