Development and Validation of a Computed Tomography–Based Radiomics Signature to Predict Response to Neoadjuvant Chemotherapy for Locally Advanced Gastric Cancer

Key Points Question Can computed tomography–based radiomics be used to predict patients’ response to neoadjuvant chemotherapy at diagnosis? Findings In this cohort study of 323 patients with locally advanced gastric cancer, no pretreatment clinicopathological characteristics were associated with response or nonresponse to treatment. However, 20 radiomic features with low intercorrelation were significantly associated with treatment response and were used to create a radiomics signature with promising clinical reliability for identifying potential responders to neoadjuvant chemotherapy. Meaning Given that no patients respond equally to neoadjuvant chemotherapy, the radiomics signature proposed in this study could guide clinicians in selecting appropriate patients for neoadjuvant chemotherapy.


Acquisition of computed tomography (CT) images
All patients underwent abdominal and pelvic contrast-enhanced CT examination within 1 week before neoadjuvant therapy at their respective hospitals. CT images were obtained from an 8-slice (GE Lightspeed Ultra 8, GE Healthcare, Hino, Japan) or a 64-slice multi-detector row CT Schering Pharma, Berlin, Germany) was injected into the patients' antecubital vein at a rate of 3.0 to 3.5 mL/s with a pump injector (Ulrich CT Plus 150, Ulrich Medical, Ulm, Germany). The arterial phase images were obtained after a 20s delay following the intravenous injection of the contrast material and portal venous phase images were obtained after 45s of delay. Both arterial and portal venous phase CT images were reconstructed with a section thickness of 5 mm. The CT images were retrieved using the picture archiving and communication system (PACS; Carestream, Canada) at each institution.

CT-images feature extraction
Tumor segmentations were performed using portal venous phase CT images as portal venous phase is better distinguishing features between the tumor mass and adjacent normal gastric tissue. Region of interest (ROI) was manually delineated along the outline of the entire visible tumor, excluding the gastric lumen, using the ITK-SNAP software (http://www.itksnap.org/).
Intraluminal air, necrotic area, vessel shadow, pericolic fat were also meticulously excluded from the contours.
After integrating the ROI of the volume of interest (VOI), the largest manually delineated ROI slice was chosen from each VOI. Before feature extraction, the ROI was resampled using a pixel size of 1 mm × 1 mm. Hounsfield units (HU) in all the images were resampled into 100 bins with HU from -200 to 600 HU, with a fixed bin size of 8 HU. Then, a 2-dimensional radiomics feature extraction was performed on the retrieved CT images using an in-house software with algorithms implemented in Matlab 2016a (Mathworks, Natick, USA). The CT images of gastric cancer (GC) lesions were resampled with a pixel size of 1mm×1mm using linear interpolation, then, separately normalized with min-max normalization to convert the pixels from -200HU to 600 HU into a range of [1, 100] of integral intensities.

Feature generation:
A series of first-order statistics, shape features and texture features were generated from the image with and without being filtered.

First-order statistics
First-order statistics describe the distribution of voxel intensities within the CT image.

3.2.Shape-based features
These features are descriptors of the two-dimensional size and shape of the tumor region.

3.3.Texture features
Textural features describe patterns or the spatial distribution of pixel intensities. They were calculated from the following matrix: When calculating texture features, the ROI was resampled using a bin number of 25. The formula of features was in https://pyradiomics.readthedocs.io/en/latest/ or 'Image biomarker standardization initiative -feature definitions' 1 .

Filter features
These features are first-order statistics and texture features extracted from filtered images. The image was filtered in the following ways, respectively: (1) Image filtration: A Laplacian of Gaussian spatial band-pass filter (∇2G) was used to derive image features at different spatial scales by turning the filter parameter between 1.0 and 2.5 (1.0, 1.5, 2.0, 2.5).
The Laplacian of Gaussian filter (∇2G) distribution is given by x and y denote the spatial coordinates of the pixel, and σ is the value of the filter parameter.

Performance of radiomics signature
In the training cohort, the area under the receiver operating characteristics (ROC) curve (AUC) was used to assess the discriminative power of the radiomics signature. The Mann-Whitney U test was also applied to evaluate the correlation between the radiomics signature and treatment response status in both the training and validation cohorts.
The C-index was calculated as a measure of the goodness of fit for the radiomics signature, equaling to the ROC curve. Calibration was used to investigate the consistency between observed outcomes and predicted values which examine how well the percentage of observed outcome matches the percentage of predicted outcome over deciles of predicted risk. This is preferably reported graphically with predicted outcome probabilities (on the x-axis) plotted against observed outcome frequencies (on the y-axis). The calibration curve was drawn by plotting ^P on the x-axis and PC = [1 + exp -(Ɣ0 + Ɣ1L)] -1 on the y-axis, where PC is the actual probability, L = logit ( ^P ), ^P is the predicted probability, Ɣ0 is the corrected intercept, and Ɣ1 is the slope estimates.

Clinical usefulness
The