Influence of inter-observer delineation variability on radiomics stability in different tumor sites

Figures & data

Figure 1. Dice coefficient and percentage of stable radiomic features. NSCLC: non-small cell lung cancer; HNSCC: head and neck squamous cell carcinoma; MPM: malignant pleural mesothelioma. Median Dice coefficient was high for NSCLC and HNSCC, whereas it was very low for MPM. The stability rate of radiomic features correlated with the median Dice coefficient.

Figure 2. Dice coefficient and percentage of stable radiomic features according to feature subgroups. NSCLC: non-small cell lung cancer; HNSCC: head and neck squamous cell cancer; MPM: malignant pleural mesothelioma. A feature is considered to be stable with an intra-class correlation coefficient >0.8.

Table 1. Percentage of stable parameters (ICC >0.8) according to calculation method.

	aGLCM (%)	mGLCM (%)	aGLRLM (%)	mGLRLM (%)
NSCLC	96	92	100	100
HNSCC	77	73	81	69
MPM	15	4	13	6

aGLCM: averaged Gray Level Co-occurrence Matrix; aGLRLM: averaged Gray Level Run Length Matrix; mGLCM: merged Gray Level Co-occurrence Matrix; mGLRLM: merged Gray Level Run Length Matrix. There are two methods for texture features definition in the two matrices. Matrices are either averaged before the texture features are computed (merged mGLCM and mGLRLM) or texture features are computed on each of 26 matrices and averaged at the end (averaged aGLCM and aGLRLM). For both matrices and all tumor sites, the method of averaging texture features at the end resulted in a slightly higher percentage of stable features.