Clustering approach to identify intratumour heterogeneity combining FDG PET and diffusion-weighted MRI in lung adenocarcinoma

Malignant tumours consist of biologically heterogeneous components; identifying and stratifying those various subregions is an important research topic. We aimed to show the effectiveness of an intratumour partitioning method using clustering to identify highly aggressive tumour subregions, determining prognosis based on pre-treatment PET and DWI in stage IV lung adenocarcinoma.

Eighteen patients who underwent both baseline PET and DWI were recruited. Pre-treatment imaging of SUV and ADC values were used to form intensity vectors within manually specified ROIs. We applied k-means clustering to intensity vectors to yield distinct subregions, then chose the subregion that best matched the criteria for high SUV and low ADC to identify tumour subregions with high aggressiveness. We stratified patients into high- and low-risk groups based on subregion volume with high aggressiveness and conducted survival analyses. This approach is referred to as the partitioning approach. For comparison, we computed tumour subregions with high aggressiveness without clustering and repeated the described procedure; this is referred to as the voxel-wise approach.

The partitioning approach led to high-risk (median SUVmax = 14.25 and median ADC = 1.26x10^-3 mm²/s) and low-risk (median SUVmax = 14.64 and median ADC = 1.09x10^-3 mm²/s) subgroups. Our partitioning approach identified significant differences in survival between high- and low-risk subgroups (hazard ratio, 4.062, 95% confidence interval, 1.21 - 13.58, p-value: 0.035). The voxel-wise approach did not identify significant differences in survival between high- and low-risk subgroups (p-value: 0.325).

Our partitioning approach identified intratumour subregions that were predictors of survival.

• Multimodal imaging of PET and DWI is useful for assessing intratumour heterogeneity. • Data-driven clustering identified subregions which might be highly aggressive for lung adenocarcinoma. • The data-driven partitioning results might be predictors of survival.