Keywords:
Thorax, Lung, Computer applications, CT, CT-Quantitative, Image manipulation / Reconstruction, Computer Applications-Detection, diagnosis, Segmentation, Technical aspects, Cancer, Tissue characterisation, Multidisciplinary cancer care
Authors:
E. Barabino1, G. Ficarra1, C. Genova1, M. Verda2, S. Casella3, S. Caprioli1, F. Grossi1, G. Cittadini1; 1Genoa/IT, 2Imperia/IT, 3Savona/IT
DOI:
10.26044/ecr2019/C-3625
Conclusion
Tang et Al [7] elaborated a radiomic signature that correlated with tumor immune microenvironment.
Our study lacks a pathological correlation but our predictive model can identify a specific radiological phenotype that is associated with a better OS in patients treated with immunotherapy for NSCLC.
Only the model with CT images filtered with a low value performed well in distinguish patients with increased OS: probably these parameters permit to markedly decrease noise,
a factor that could easily generate a so-called “pseudotexture”,
without losing information about texture.
Machine Learning is frequently compared to a black box [8] due to the impossibility of truly understanding how it works or to achieve insights about how it “thinks”.
The limited number of patients and dimensionality of textural features are the main limit of our study a problem well-known as “large p,
small n”.
PCA was essential to perform features selection particularly in our study where texture features are highly correlated because derived from co-occurrence matrices,
therefore,
the same features is calculated multiple times with different angle and distance.
Moreover,
our cohort is exclusively composed of patients that were highly treated with multiple lines of chemotherapy and radiotherapy while immunotherapy has a higher response rate in chemo-naive patients.
A larger cohort of patients can create more reliable predictive models that can permit subgroup analysis.
Finally,
blood-based biomarkers or histological samples could be unreliable in predicting response to immunotherapy however they can be used in machine learning algorithm combined with textural features to refine classification.
In conclusion,
Machine Learning and Texture Analysis are valuable tools in the prediction of response to immunotherapy in NSCLC but research is needed to explore in clinical decision-making.