Radiomic Nomogram: Pretreatment Evaluation of Local Recurrence in Nasopharyngeal Carcinoma based on MR Imaging

Background: To develop and validate a radiomic nomogram incorporating radiomic features with clinical variables for individual local recurrence risk assessment in nasopharyngeal carcinoma (NPC) patients before initial treatment. Methods: One hundred and forty patients were randomly divided into a training cohort (n = 80) and a validation cohort (n = 60). A total of 970 radiomic features were extracted from pretreatment magnetic resonance (MR) images of NPC patients from May 2007 to December 2013. Univariate and multivariate analyses were used for selecting radiomic features associated with local recurrence, and multivariate analyses was used for building radiomic nomogram. Results: Eight contrast-enhanced T1-weighted (CET1-w) image features and seven T2-weighted (T2-w) image features were selected to build a Cox proportional hazard model in the training cohort, respectively. The radiomic nomogram, which combined radiomic features and multiple clinical variables, had a good evaluation ability (C-index: 0.74 [95% CI: 0.58, 0.85]) in the validation cohort. The radiomic nomogram successfully categorized those patients into low- and high-risk groups with significant differences in the rate of local recurrence-free survival (P <0.05). Conclusions: This study demonstrates that MR imaging-based radiomics can be used as an aid tool for the evaluation of local recurrence, in order to develop tailored treatment targeting specific characteristics of individual patients.

slowly or periodically. Our textural features mainly consisted of the gray-level co-occurrence matrix (GLCM) and the gray-level run-length texture matrix (GLRLM).
The GLCM is a pixel matrix function of the distance and angle, which quantifies the correlation by calculating a certain distance and a certain direction between two gray matrices; in this manner, the matrix reflects the integrated information in the direction, interval, amplitude, and frequency [1]. The GLRLM quantifies gray level runs in an image [2,3]. A gray level run is defined as the length number of continuous pixels that have the same gray level value. We extracted 22 features from the GLCM and 14 features from the GLRLM [4]. The radiomic features in the GLCM mainly comprised energy, entropy, correlation, contrast, homogeneity, autocorrelation, mean, variance, dissimilarity, and angular second moment. The radiomic features in the GLRLM mainly comprised run length non-uniformity, short/long run emphasis, and gray level non-uniformity.
Wavelet features: In our study, the undecimated three-dimensional wavelet transform was used to decompose the original image. Consider L and H to be low-pass and high-pass functions, X to be the decomposing image, and the wavelet decompositions of X to be labeled as , , , We can obtain eight new images that are decomposed in three directions (x, y, z), where the size of each decomposition is equal to that of the original image and each decomposition is shift invariant. For each decomposition, we computed the first-order statistics and textural features described above. This resulted in 424 features.
Ultimately, we extracted 485 features from each of the series.
In conclusion, 970 radiomic features were extracted from the MR images in our study.

Feature selection
In this paper, we selected the patients who experienced recurrence within 1 year and the patients who did not experience any recurrence for >5 years, as a 0/1 label for the selection of the features. We used the recursive feature elimination with logistic regression algorithm (LR-RFE) to select the features, and determined the features of the model based on the highest AUC value of the classification results [5,6]. From the CETI-w images, we selected eight features as follows: CET1-w_3_fos_median, As demonstrated, our features were all from images that were decomposed by the undecimated three-dimensional wavelet transform. Numbers from '1' to '8' were used to mark the different respective forms of the wavelet transform. The eight features of the CET1-w images were composed of (marked by "1"), (marked by "3"), (marked by "4"), and (marked by "6") images. Similarly, the seven features of the T2-w images were composed of (marked by "1"), (marked by "4"), (marked by "6"), and (marked by "7") images.
In the later discussion, we let X reflect the three-dimensional image matrix with N voxels used to analyze the first-order statistics features and the shape-and size-based features that were selected by the LR-RFE algorithm. Meanwhile, we consider the GLCM and GLRLM to be a matrix size , defined as . Here, the element represents the number of times the combination of intensity levels occurs in two pixels in the images that are separated by a distance of δ pixels in direction α, and is the number of discrete gray level intensities, and is the number of different run lengths. In addition, let separately refer to the marginal row and column probabilities. Their detailed explanation is as follows:

CET1-w_3_fos_median:
The first-order statistics feature that describes the median value of the intensity levels in CET1-w images in images.

CET1-w_Surface_to_volume_ratio:
surface to volume ratio = CET1-w_3_fos_mean: The first-order statistics feature that describes the mean value of the intensity levels in CET1-w images in images.