Automatic visual-like classification of lung tumour heterogeneity in DCE-CT sequences

Purpose: Tumour heterogeneity is an important prognostic factor, as high intra-tumour heterogeneity showed to be associated with higher tumour grades. However, its assessment is still mostly accomplished subjectively through visual procedure. This work presents an automatic approach to classify the heterogeneity levels in lung tumour as performed through visual analysis. Methods and Materials: 40 datasets referring to 13 patients (age range 36-81 years) with NSCLC, who underwent axial DCE-CT, were considered. Two 25-year experienced Readers chose the most representative slices in the DCE-CT sequences, outlined each lesion and its most significant regions. Then, each slice was assigned a class, according to a proper taxonomy for heterogeneity levels previously defined: homogeneous, macro-inhomogeneous (i.e., different homogeneities together), and micro-inhomogeneous. A statistical voxel-based index was devised to quantify the heterogeneity, then represented in colorimetric maps. The values were grouped into regions subsequently compared with those drawn by radiologists. Results: Results for the three classes were computed in terms of specificity (SP) and sensitivity (SE). Our approach proved to be extremely specific, mostly for homogeneous (SE=77%, SP=93%) and macro-inhomogeneous (SE=75%, SP=90%) tissues. On the other hand, the most indefinite micro-inhomogeneous tissue also shows a high specificity (SE=86%, SP=87%). Conclusion: The approach developed allows an automatic classification of heterogeneities, with a reduction of both intra- and inter-observer variability. This represents a novel approach acting as a second radiologist in the heterogeneity assessment, which could yield a great benefit for patient stratification and constitutes a valid tool to assist radiologists in daily clinical activities.