Supplement for the manuscript “Classification of early-stage non-small cell lung cancers in CT images into histological types using radiomic features: Interobserver delineation variability analysis”
1Semi-automatic contouring using Pinnacle3 model-based segmentation
In the present radiomics study, we used a volume-of-interest (VOI) contoured by semi-automatic segmentation in addition to the manually delineated ones. Here, we summarize our procedure for semi-automatic delineation, using Pinnacle3 v9.10 (Philips). After setting the window and level at 1700 and -300, respectively, a “Sphere” is added as a region-of-interest (ROI) from the organ model library (Fig. S-1-(a)), and then modified by adjusting the scales in all three dimensions (Fig. S-1-(b)). In the organ model settings window (Fig. S-1-(c)), we set 500 for “Min CT Value”, 1500 for “Max CT Value”, 80 for “Max Gradient CT#/mm”, “Greater than” for “Avoid CT Values”, 1500 for “Threshold Value”, “Yes” for “Settings acceptable for adaptations”, and the defaults for the rest of the parameters.
With this organ model setting,the automated adaptation is applied 1-3 times. Figure S-1-(d) shows the result of the first adaptation, while Fig. S-1-(e) shows the result of the sequential second adaptation. In the present study, the automated ROI was judged to be accurate for shape, but not accurate for size. Therefore, a modification was made by enlarging the ROI in Fig. S-1-(e) isotropically.
2Feature extraction procedure
In the present radiomics study, we calculated 8 shape/size features, 10 non-texture global features, and 42 texture features, as shown in Table. 2 of the article. Whereas the shape/size features depend only on the shape and size of VOI, the other features depend on the intensity and/or its distribution in theanalyzed images. The latter, therefore has different values among the three-dimensional (3D) wavelet decomposed images and the original CT image. Figure S-2 shows anexample of the decomposed images as well as the corresponding original CT image. The 3D wavelet transform is referred to in the supplement in Ref. [H.J.W.L. Aerts, et. al., Nature communications, 5; (2015)]. Inthe decomposed images, the range of image intensity is broader than in the original CT image. In this study, we found that the intensity range in all images fromall patients was [-3025, 2605]. Following Ref. [H.J.W.L. Aerts, et. al., Nature communications, 5; (2015)], this study employed 25-Hounsfield unit equal width bins (which means 225-bin number for above range) in the texture feature evaluation.
3Validation using the Cancer Imaging Archive (TCIA)
In this supplement, we present other validation results using an open access database, called TCIA.
TCIA is a service thatde-identifies and hosts a large archive of medical images of cancer, accessible for public download. The data are organized as “Collections”, typically patients related by a common disease (e.g. lung cancer), image modality (MRI, CT, etc.), or research focus. DICOM is the primary file format used by TCIA for image storage. Supporting data related to the images such as patient outcomes, treatment details, genomics, pathology, and expert analyses are also provided when available.Forthe validation of our model, the “NSCLC-Radiomics” datafromTCIA was used. In “NSCLC-Radiomics”, 29 early-stage GTV-completed datasets were selected. Among them, adenocarcinoma and squamous cell carcinoma were diagnosed in 15 and 14 cases, respectively.The features were then extracted in the way indicated in the manuscript and their values were centerized[A1]. Table S-1 shows the AUC valuesfor theTCIA dataset, where the feature space used was“Three-VOI Non-Corr. 8”,and the model parameters werelearned with our hospital data (4 ROI’s). The resulting AUC result is high, meaning that the radiomics predictability of the present model is credible.
Model / AUCOriginal GTV / 0.768
Semi-auto / 0.644
Oncologist 1 / 0.548
Oncologist 2 / 0.720
[A1]This word is wrong, but I don’t understand the intended meaning to offer a suitable revision. Please clarify so that I can offer a better alternative.