Assessment of Intratumoral and Peritumoral Computed Tomography Radiomics for Predicting Pathological Complete Response to Neoadjuvant Chemoradiation in Patients With Esophageal Squamous Cell Carcinoma

Key Points Question Are peritumoral radiomics features extracted from pretreatment computed tomography images predictive of pathological complete response following neoadjuvant chemoradiation in patients with esophageal squamous cell carcinoma? Findings In this diagnostic study of 231 patients, the developed model integrating intratumoral and peritumoral radiomics features achieved improvement of predictive performance (area under the receiver operating characteristic curve, 0.852) compared with the conventionally constructed model merely using intratumoral radiomics features (area under the receiver operating characteristic curve, 0.730). Meaning Peritumoral radiomics may provide additional predictive value for treatment response estimation in esophageal squamous cell carcinoma and thus benefit individualized therapeutic strategies.


Radiomics Feature extraction and definitions
The resampled voxel sizes were set to 1×1×5 mm³ voxels to standardize the slice thickness. Image intensities were binned by 25 HU and voxel array shift was set on 1000. Care was taken to threshold the Hounsfield units (HU) of the CT scan (-100~400 HU) in order to remove air and bone pixels.
Original images depicted the baseline structural characteristics. Wavelet filtration (high and low pass filters) filtered original images directionally with x, y and z directions respectively, yielding 8 different combinations of decompositions. LoG filter, also called edge enhancement filter, convolved the image with the second derivative (Laplacian) of a Gaussian kernel. The Gaussian kernel is sensitive to areas with rapidly changing intensities and enhancing edges. The width of the filter in the Gaussian kernel is determined by and we adopted with the range of 1-5 mm at 1 mm interval to achieve both fine (low values) and coarse (high values) textures. These filtrations could capture more detailed disruptions in different orientations.
The features can be divided into 3 groups: (a) first-order statistics, describing commonly used and basic metrics for distributions of the voxel intensity within the ROI; (b) shape features, including descriptors of the two-dimensional and three-dimensional morphological properties (size and shape) of the ROI; and (c) second-order features, consisting of high-dimensional textual features quantifying the spatial distribution of pixel intensities in all three-dimensional directions, which takes the spatial distribution of each voxel and the neighboring voxels into consideration. Most features defined below were in accordance with feature definitions as described by the Imaging Biomarker Standardization Initiative (IBSI), which were available in a separate document by Zwanenburg et al. (2016). 1 Here, we show the description of the radiomic features used in our optimal combined model:

First order features
First-order statistics describe the distribution of voxel intensities within the image region defined by the mask through commonly used and basic metrics.
Let: ▪ be a set of voxels included in the ROI ▪ ( ) be the first order histogram with discrete intensity levels, where is the number of non-zero bins, equally spaced from 0 with a width defined in the bin width parameter ▪ ( ) be the normalized first order histogram and equal to ( )/

1) Median
The median gray level intensity within the ROI.

2) Kurtosis
Where 4 is the 4 th central moment.
Kurtosis is a measure of the 'peakedness' of the distribution of values in the image ROI. A higher kurtosis implies that the mass of the distribution is concentrated towards the tail(s) rather than towards the mean. A lower kurtosis implies the reverse: that the mass of the distribution is concentrated towards a spike near the Mean value.

Gray Level Size Zone Matrix (GLSZM) Features
A Gray Level Size Zone (GLSZM) quantifies gray level zones in an image. A gray level zone is defined as the number of connected voxels that share the same gray level intensity. A voxel is considered connected if the distance is 1 according to the infinity norm (26-connected region in a 3D, 8-connected region in 2D). In a gray level size zone matrix ( , ) the ( , ) th element equals the number of zones with gray level and size appear in image. Contrary to GLCM and GLRLM, the GLSZM is rotation independent, with only one matrix calculated for all directions in the ROI. 2 As a twodimensional example, consider the following 5x5 image, with 5 discrete gray levels: The GLSZM then becomes:

3) Inverse Variance
Note that =0 is skipped, as this would result in a division by 0.

4) Cluster Shade
Cluster Shade is a measure of the skewness and uniformity of the GLCM. A higher cluster shade implies greater asymmetry about the mean.

RNA sequencing procedure
Total format. In order to gain clean data, reads that included adapters, contained high proportion of N for no base information, or had low quality were removed. Clean reads were aligned to the hg19 human genome using TopHat v2.1.1, and mapped reads were assembled for each sample by Cufflinks v2.
Annotation was performed according to the RefSeq database for mRNA and the GENCODE v21 database for lncRNA. To assess expression level, read counts were normalized and calculated as fragments per kilobase of exon per million fragments (FPKM) by Cuffdiff v2.2.1.

R packages
Radiomics features were harmonized to reduce the multicenter effect caused by different scanner and protocol settings. According to the statistical distribution of the dataset, nonparametric form of the model was adopted in which ComBat determined the transformation for each feature separately using "sva" R package. 4 Feature robustness was tested by intraclass correlation coefficients (ICCs) using "irr" R package. 5 The AUC of the receiver operating characteristic (ROC) curve was calculated and compare by "pRCO" R package. 6 The raw genomic data was preprocessed (background correction, log2-transformation and quantile normalization) using the Bioconductor package "affy". 7 eFigure 1.