Nuclear morphometry features distinguish cell types and outcome groupings in lung cancer


Session type:

Katey Enfield1,Spencer Martin2,Erin Marshall2,Zhaolin Xu3,Martial Guillaud2,Calum MacAulay2,Wan Lam2
1British Columbia Cancer Research Centre,2BC Cancer,3Dalhousie University



The shape and organization of tumour cell nuclei has been shown to be associated with aggressiveness in several cancers, including prostate and cervical cancer. Fortunately, this data can be obtained using a single quantitative nuclear stain on FFPE tissues. We sought to determine whether nuclear features could be utilized to distinguish cell types within the lung tumour microenvironment and to classify outcome groupings.


A tissue microarray of non-small cell lung cancer samples was stained with the quantitative nuclear dye, Feulgen/thionin. The stained slides were scanned using a hyperspectral imaging platform, images were spectrally unmixed, and cell nuclei were identified using a segmentation algorithm. Nuclei were then analyzed to identify 245 nuclear morphometry features. Nuclear features were assessed in order to dichotomize good outcome (alive at 5 years) and bad outcome (survival of 3 years or less) patients, as well as several cell types within the tumour microenvironment. Finally, we assessed the spatial organization of these different classes of nuclei in good and poor outcome cases.


Nuclear features derived from thionin stain classified epithelial, stromal, and immune cells. Of the 245 nuclear features assessed, 87 differed significantly between good and poor outcome groups. While many distinguishing features were identified in tumour cell nuclei, we were also able to identify distinguishing features of non-tumour cell nuclei. Using pairs of features, good and poor outcome cases were stratified with up to 69% accuracy. However, the addition of spatial information to our classifier resulted in the improvement of sample stratification to 86% accuracy.


Using a single nuclear stain on FFPE lung cancer histology specimens, we were not only able to sub-classify cell types, but also to distinguish between good and poor outcome cases. The spatial organization of these cell types greatly improved our classifier, indicating the potential of this approach for prognostic purposes.