Table 1.
Baseline characteristics of the clinical cohort.
Table 2.
Radiomic features extracted from contrast and non-contrast enhanced CT images.
GLCM, grey level co-occurrence matrix; GLRLM, grey level run length matrix; GLSZM, grey level size zone matrix; GLDZM, grey level distance zone matrix; NGTDM, neighbourhood grey tone difference matrix.
Fig 1.
Workflow used for selecting stable radiomic features.
The workflow was repeated three times using three different groups (a “mixed group” with contrast and non-contrast enhanced CT images, a “contrast group” with contrast enhanced CT images and a “non-contrast group” with non-contrast enhanced CT images) as input data. The input group was divided in three sub-groups (n = 71 for the mixed group, n = 46 for the contrast group and n = 25 for the non-contrast group) and processed to extract the features. Stable features, identified as the ones with similar distributions among the sub- groups, were further analysed. Feature with different distributions among sub-group (identified as unstable) were not further investigated.
Fig 2.
Stable and unstable radiomic features extracted from the three groups.
GLCM, grey level co-occurrence matrix; GLRLM, grey level run length matrix; GLSZM, grey level size zone matrix; GLDZM, grey level distance zone matrix; NGTDM, neighbourhood grey tone difference matrix; *, feature computed with merging.
Fig 3.
Stability of the radiomic features for the three groups assessed.
Features were divided in 2D or in 3D.
Table 3.
List of stable radiomic features not statistically different if extracted using one segment layer at a time or considering the whole tumour volume in CT images of the oesophagus.
In bold, common features among three cohorts considered that showed to be stable and dimensionality and contrast agent independent. GLCM, grey level co-occurrence matrix; GLRLM, grey level run length matrix; GLSZM, grey level size zone matrix; GLDZM, grey level distance zone matrix; NGTDM, neighbourhood grey tone difference matrix; *, feature computed with merging.
Fig 4.
Kaplan-Meier survival curves of the cohort used in this work.
Patients divided in a) low-risk group (zone distance varianceGLDZM from 0.15 to 0.85), intermediate-risk group (zone distance varianceGLDZM from 0.86 to 1.69) and high-risk group (zone distance varianceGLDZM from 1.70 to 6.2) based on the prognostic score and b) group 1 (zone distance varianceGLDZM < 1.70) and group 2 (zone distance varianceGLDZM = > 1.70) considering a threshold of zone distance varianceGLDZM. There was a significant difference in OS when dividing in low, intermediate and high-risk group (X2 = 7.37, df = 1, p-value = 0.007) or when using a threshold of zone distance varianceGLDZM (X2 = 7.692, df = 1, p-value = 0.006).