Radiomics Analysis and Unsupervised Self-Organizing-Map Technique to Predict Radiation-Induced Pneumonitis in Patients with Lung Cancer
Bagher-Ebadian H, Wu Q, Ghanem A, Brown S, Ajlouni M, Simoff M, Movsas B, and Chetty I. Radiomics Analysis and Unsupervised Self-Organizing-Map Technique to Predict Radiation-Induced Pneumonitis in Patients with Lung Cancer. International Journal of Radiation Oncology Biology Physics 2020; 108(2):E49-E50.
International Journal of Radiation Oncology, Biology & Physics
Background: Lung radiation pneumonitis (RP) is one of the major toxicities experienced by lung cancer patients (13%–37%) receiving thoracic radiation therapy (RT) (Kocak Z, et al. Int. J. Radiat. Oncol. Biol. Phys. 2005 and Rodrigues G, et al. Oncol. 2004). Investigation of imaging biomarkers that can predict the incidence of RP can be useful toward reduce the probability of RP development. In this pilot study, we investigated the application of Kohonen Self-Organizing Map (K-SOM; a type of adaptive model that is trained using unsupervised competitive-learning (Kohonen T, Springer 1995)), for analysis of radiomic features extracted from normal lung tissue of non-RP and RP patients. The goal was to build an adaptive model to stratify patients and to reveal dominant characteristics of radiomics features for patient stratification. These characteristics are likely to remain unnoticed using conventional statistical analysis.
Objectives: To perform deep and unsupervised analysis of radiomics features extracted from planning CT images of normal lung tissue for patients with locally advanced, non-small cell lung cancers (NSCLC) to characterize and stratify patients with and without radiation-induced RP.
Methods: Planning CT images of 41 patients (14 with RP and 27 with no evidence of RP) with stage-III lung cancers, treated with IMRT/3D-CRT, were studied. One hundred sixty eight radiomics features were extracted from the volume of normal lung tissue receiving ≥20 Gy, excluding the ITV, according to the following 8 different classes: Intensity Histogram Based Features (IHBF), Gray Level Run Length (GLRL), Law’s Textural information (LAWS), Discrete Orthonormal Stockwell Transform (DOST), Local Binary Pattern (LBP), 2D-Wavelet Transform (2DWT), 2D-Gabor Filter (2DGF), and Gray Level Co-Occurrence Matrix (GLCM). A K-SOM ([9×9]-neurons) was constructed using 168 radiomic features and was trained and evaluated using random-permutation-sampling method (100-iterations, 67% and 33% for modeling and testing respectively). Regions pertaining to the dominant characteristics of the discriminant features were identified in the K-SOM space and used to calculate stratification power of the classifier.
Results: The K-SOM revealed 3 dominant regions associated with the RP status: high-certainty-non-RP, high-certainty-RP, and intermediate/uncertain region. The average of AUC, positive predictive and negative predictive values were 76.20%, %71.65, and 82.12% respectively. Correlation-based dissimilarity analysis (r<0.2, p-value<0.05) ranked and revealed the stability and robustness of the 6 discriminant features as follows: Skewness, Entropy, and Moment (3 and 4) from IHBF, Entropy-CA1 from DOST, and Short-Run-Emphasis from GLRL.
Conclusions: The results of this pilot study, albeit subject to confirmation in a larger patient population, suggest a potential role for the use of an unsupervised method for radiomics-based stratification and prediction of radiation-induced pneumonitis in patients with locally advanced NSCLC.