4119
Semi-automatic ROI registration and texture feature extraction for pre- and post- neoadjuvant chemotherapy contrast-enhanced MR in predicting response of mass-like breast cancers1Peking University Cancer Hospital & Institute, Beijing, China
Synopsis
In mass-like breast cancers, semi-automatic ROI registration from pre- to post-treatment MR images, and combined with usage of texture feature extraction for automatically segmentation of enhanced areas, the ability for predicting residual tumors after neoadjuvant chemotherapy can be improved significantly and cost lesser time in imaging reading.
INTRODUCTION
Achieving complete response after neoadjuvant chemotherapy(NAC) for breast cancers indicates favorable prognosis. Texture analysis can provide information of intra-tumoral heterogeneity, which is also a prognostic factor. However, it takes time to locate the exact live areas at the site of original tumor bed after NAC and extract texture features. A method is used in this study to automatically register the manually drawn regions of interests (ROI) on pre-NAC MR to post-NAC MR, and then do the segmentation within in RIOs to achieve suspicious residual tumor areas, and then make texture analysis. The study was made on contrast enhanced MR images to investigate the clinical application of such method in predicting complete response of mass-like lesions after NAC. In mass-like breast cancers, semi-automatic ROI registration from pre- to post-treatment MR images, and combined with usage of texture feature extraction for automatically segmentation of enhanced areas, the ability for predicting residual tumors after neoadjuvant chemotherapy can be improved significantly and cost lesser time in imaging reading.METHODS
Consecutive breast cancer patients who received NAC before operation were enrolled. MR images both before and after NAC of those mass-like lesions on report were retrieved to make subtraction images of early and late phases to pre-enhanced phase on MR DCE sequences. Regions of interest were drawn manually on early phase of pre-NAC images and covered the whole enhanced areas. After automatic registration of post-NAC images to pre-NAC images using 3D slicer software, volumetric ROIs were copied directly to post-NAC images and automatic threshold segmentation were used to pick only voxels which had enhancement above the set threshold (fig 1). Texture features were extracted using a house made radiomics software also based on 3D slicer, with 44 features acquired (including volume), so totally 174 features were generated from pre- and post-NAC MR and early- and late- enhanced phases. Operational pathology reports were retrieved to divide the patients into residual and non-residual groups. Comparisons were made between groups, and Cox regression analysis were used to pick the valuable parameters and assess the diagnostic abilities.RESULTS
Totally 52 mass-like breast cancer were enrolled, with 32 pathologically found invasive cancers (residual tumors) and 20 complete response either with non-cancer cell or DCIS (non-residual tumor). There were 13 features found to be significantly different between groups after primary univariate analysis group (P≤0.05). Multivariate Cox regression analysis found 5 parameters to be independent factors (including volume, kurtosis and energy of pre-NAC in early phase, SImax in early phase, kurtosis of post-NAC in late phase), and yield a diagnostic accuracy of 98.1%, 100% sensitivity and 96.8% specificity in predicting residual cancer.DISCUSSION
To predict complete response after NAC can help the clinicians to avoid overtreatment. MR had the most accuracy in predicting pathological complete response by 89% to 92% in literature review [1], and is hard to improve by current reading methods. Proper registration of pre-treatment lesion ROI to post-treatment images can help the radiologists to found original tumor bed, and using threshold segmentation can be easier and more time-saving in quantitatively picking the suspicious areas comparing with human judgment by bare eyes.
Volume is always a well accepted prognostic factor that relating with treatment effect, but may only in mass-like lesions as in this study. Non-mass like lesions are usually large and difficult to locate and cover completely, thus was discarded in out study.
Though early phase of contrast enhanced images were considered to be best in depicting cancer areas, and highest enhancement do showed significance as in other studies [2]. However, heterogeneity in late enhance phases, represented in texture analysis by numerous features, is found can be signified by kurtosis of signal intensity, and is an independent factor with high significance.
CONCLUSION
In mass-like breast cancers, semi-automatic ROI registration from pre- to post-NAC MR images, and combined with usage of texture feature extraction for automatically segmentation of enhanced areas, the ability for predicting residual tumors after neoadjuvant chemotherapy can be improved significantly and cost lesser time in imaging reading.Acknowledgements
Thanks for Hui Liu, engineer, for developing the radiomics software on 3D slicer platform.References
[1] Yuan Y, Chen XS, Liu SY, et al. Accuracy of MRI in prediction of pathologic complete remission in breast cancer after preoperative therapy: a meta-analysis. AJR Am J Roentgenol 2010, 195(1):260-268.
[2] Mori N, Pineda FD, Tsuchiya K, et al. Fast Temporal Resolution Dynamic Contrast-Enhanced MRI: Histogram Analysis Versus Visual Analysis for Differentiating Benign and Malignant Breast Lesions. AJR Am J Roentgenol 2018, 211(4):933-939.
Figures
