Random forest analysis to predict disease-free survival using FDG-PET and CT in non-small cell lung cancer

Congress:

ECR 2018

Poster Number:

C-2980

Type:

Scientific Exhibit

Keywords:

Hybrid Imaging, Lung, Computer applications, PET-CT, Image manipulation / Reconstruction, Neural networks, Computer Applications-Detection, diagnosis, Outcomes analysis, Experimental investigations, Tissue characterisation, Cancer, Outcomes

Authors:

M. Kirienko¹, L. Lozza², N. Gennaro¹, A. Rossi¹, E. Voulaz¹, A. Chiti¹, M. Sollini¹; ¹Milan/IT, ²Bergamo/IT

DOI:

10.1594/ecr2018/C-2980

DOI-Link:

https://dx.doi.org/10.1594/ecr2018/C-2980

Fig. 2: The importance grade of the features.

Fig. 3: Graphical representation of patients belonging to the training set, depending...

Fig. 1: Disease-free survival prediction results for the PET+CT data. Kaplan-Meier...

Fig. 4: Fig. 4 Low-risk patient:CT, axial PET, PET/CT and a coronal reconstruction of...

Aims and objectives

We previously identified a radiomic signature capable of predicting disease-free survival (DFS) in NSLC patients undergoing surgery1. In this study, we evaluated the same population with a non-parametric, multivariate analysis using a random forest model, aimed at predicting DFS from a combination of input variables.

Methods and materials

Random forests for classification were developed keeping the same training and validation sets as for the parametric analysis, to predict DFS. Seven different combinations of variables were considered: Clinical (263 patients, 5 features), CT (295, 41), PET (258, 43), PET+CT (258, 84), CT+Clinical (263, 46), PET+Clinical (231, 48), PET+CT+Clinical (231, 89) Random forests for classification were developed keeping the same training and validation sets as for the parametric analysis. The outcome to be predicted was the DFS considered until the date of last access or...

Results

The highest AUC obtained on the validation set was 0.79. The corresponding model was a forest with 10000 trees, 6 as split dimension, 0.25/0.75 as relative weight and on the dataset containing CT and clinical features. The dataset was the one with only the feature with importance greater than the 80th quantile. In particular the nine variable se-lected were: ’Contrast_NGLDM’, ‘Correlation_GLCM', ‘EntropyH', ‘LZHGE_GLZLM', ‘SZHGE_GLZLM', ‘Sphericity', ‘maxValue', ‘minValue', ‘stdValue'. The importance grade of the feature isillustrated in Figure 2. The graphical representation of patients belonging to...

Conclusion

Innovative statistics analysis are a promising tool to select robust radiomics signatures

Personal information

References

1. Kirienko, M., Cozzi, L., Antunovic, L. et al. Eur J Nucl Med Mol Imaging (2018) 45: 207. https://doi.org/10.1007/s00259-017-3837-7