Item 9 — Data pre-processing

Describe preprocessing steps to allow other investigators to reproduce them. Specify the use of normalization, resampling of image size, change in bit depth, and/or adjustment of window/level settings. If applicable, state whether the data have been rescaled, threshold-limited (“binarized”), and/or standardized. Specify processes used to address regional formatting, manual input, inconsistent data, missing data, incorrect data type, file manipulations, and missing anonymization. State any criteria used to remove outliers…. When applicable, include description for libraries, software (including manufacturer name and location and version numbers), and all option and configuration settings.

Explanation and Elaboration: Image preprocessing involves transforming raw imaging or clinical data into a format suitable for model input. Preprocessing significantly affects model performance and reproducibility. It is not sufficient to generally describe the types of preprocessing steps that were undertaken; each step must be described in complete detail so that someone else could exactly reproduce the preprocessing procedure. This often requires listing the utilities and libraries and their versions, along with command line switches, argument values and parameter settings used. When custom software is written to perform preprocessing, the source code of that software should be made available, such as in a supplemental section.

Examples
Reports were segmented into sections—clinical context, protocol, results, and conclusion—with Python, version 3.10 (Python Software Foundation). The report text itself was left unaltered. Reports were initially written in French and were not translated to English for this study. For reference, texts shown in Figure 1 and Tables 1, S1, and S2 were translated to English by the authors and manually modified for de-identification. [Materials and methods section (24), CC-BY 4.0]
Perfusion image series underwent motion and surface coil inhomogeneity correction. Motion correction utilized nonrigid image registration in an iterative manner. To compensate for substantial image contrast variation during the contrast bolus passage, instead of directly registering perfusion images against each other, synthetic perfusion series were derived from a Karhunen‐Loève transform. Motion correction was achieved by registering perfusion images pairwise with the synthetic series. The detailed algorithm was presented in Xue et al…. After correcting respiratory motion, surface coil inhomogeneity was corrected using the proton density images and the normalized intensities were converted to gadolinium concentration units (mmol/L)…. To compensate for heart rate variation and mis-triggering, the perfusion series was temporally resampled using linear interpolation which also compensated for possible missed triggers. This interpolation resulted in a fixed sampling corresponding to a heart rate of 120 bpm. The temporal resampling step did not lead to spatial blurring since it was performed after motion correction. Since the gadolinium concentration series was corrected for signal nonlinearity and surface coil inhomogeneity, it had the benefit of reducing the dynamic range and providing a fixed signal range for neural nets, compared with perfusion intensity images. This image series was spatially upsampled to 1.0 mm2 spatial resolution and the central field of view (176 × 176 mm2) was cropped. The left ventricular (LV) blood pool was detected from the arterial input function series which was imaged at the basal plane at diastole. The location of the LV blood pool from this step was used to center the cropped image…. For the short axis perfusion slices, the LV endo-and epicardial boundaries were manually traced, together with the right ventricle (RV) (Fig 1). The RV insertion point was determined from the segmented right ventricular and LV center as the rightmost pixel. The training and test datasets were carefully labeled by one operator (10 years of experience in perfusion imaging). [Supplementary material E1 (25), CC-BY 4.0]
All the CT images were converted from Digital Imaging and Communications in Medicine format to NRRD format using the Medical Imaging Interaction Toolkit…. The CT images contained four to six mice each; thus, the first step was to crop each CT scan to obtain CT image data from a single mouse. Coordinates of the lung edges were identified from the reference mask and were used to find the x, y, and z coordinates to crop the corresponding CT image. This cropping resulted in a total of 1654 mice in group A (1500 mice for training and validation and 154 for testing) and 237 mice in group B. Cropped images were resized to achieve an isotropic resolution of 0.35 mm. The reference lung masks (ground truth) were segmented semi automatically, as described by Zhou and colleagues … for both fibrotic and normal lungs. Briefly, each CT was thresholded between −900 HU and −100 HU to obtain an approximate lung segmentation. Subsequently, incorrectly assigned voxels, mostly in the trachea and primary bronchi areas, were manually removed. The CT images from group C contained one or two mice each. Images with two mice were cropped as described previously. The total number of mice in group C was 20 (16 used to retrain the model, and four used for testing). Each cropped CT image was then resampled to an isotropic resolution of 0.2 mm. Ground truth lung masks were again segmented semi automatically. Finally, given the two-dimensional architecture of the proposed method, each axial section in the three-dimensional images from all groups was extracted, and the intensity was normalized between zero and 1. [Methods (26), CC-BY 4.0]