R Kakino^1,2,3*, M Nakamura^1,2, T Mitsuyoshi^2,4, T Shintani², M Kokubo⁴, Y Negoro⁵, M Fushiki⁶, M Ogura⁷, S Itasaka⁸, C Yamauchi⁹, S Otsu¹⁰, T Sakamoto¹¹, M Sakamoto¹², N Araki¹³, H Hirashima², T Adachi^1,2, Y Matsuo², T Mizowaki², (1) Division of Medical Physics, Department of Information Technology and Medical Engineering, Human Health Sciences, Graduate School of Medicine, Kyoto University, (2) Department of Radiation Oncology and Image-applied Therapy, Graduate School of Medicine, Kyoto University, (3) Research Fellow at Japan Society for the Promotion of Science, (4) Department of Radiation Oncology, Kobe City Medical Center General Hospital, (5) Department of Radiology, Tenri Hospital, (6) Department of Radiation Oncology, Nagahama City Hospital, Nagahama, (7) Department of Radiation Oncology, Kishiwada City Hospital, Kishiwada , (8) Department of Radiation Oncology, Kurashiki Central Hospital, (9) Department of Radiation Oncology, Shiga General Hospital, (10) Department of Radiation Oncology, Kyoto City Hospital, (11) Department of Radiation Oncology, Kyoto-Katsura Hospital, (12) Department of Radiology, Japanese Red Cross Fukui Hospital, (13) Department of Radiation Oncology, National Hospital Organization Kyoto Medical Center

• 
  R Kakino

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: To predict local recurrence (LR) and distant metastasis (DM) for early stage non-small cell lung cancer (NSCLC) patients after SBRT in multi-institution by radiomic features extracted from breath-hold CT images.

Methods: A total of 573 patients who underwent SBRT between January 2006 and March 2016 met eligibly criteria were included in the study. Patients were dichotomized into two datasets: training (464 patients in ten institutions) and test (109 patients in one institution). A total of 944 radiomic features (original: 107, Laplacian of Gaussian: 465, wavelet: 372) were extracted from manually segmented gross tumor volumes (GTVs).. Feature selection was performed by analyzing inter-rater segmentation reproducibility, GTV correlation, and inter-feature redundancy. Important features were then selected by adaptive least absolute shrinkage and selection operator. Nine clinical factors such as histology and GTV size were also utilized. Three prognostic models (clinical, radiomic, and combined models) for LR and DM were constructed using random survival forest (RSF) dealing with total death as a competing risk in the training dataset. Robust models with the optimal hyper-parameters were determined using 5-fold cross validation. Final model was validated by concordance index at three years and risk score-based stratification in the test dataset.

Results: The concordance indices at 3 years of clinical, radiomic, and combined models for LR were 0.57, 0.55, and 0.61, respectively, whereas those for DM were 0.59, 0.67, and 0.68, respectively, in the test dataset. The combined model for DM significantly discriminated its cumulative incidence (p < 0.05). The variable importance of RSF in the combined model for DM indicated that two radiomic features (wavelet.LH_glcm_MCC, wavelet.LL_glcm_imc2) were more important than other clinical factors.

Conclusion: The radiomics approach with RSF for competing risks using breath-hold CT-based radiomic features can predict DM in early-stage NSCLC patients who underwent SBRT.

Keywords

CT, Lung, Feature Extraction

Taxonomy

IM- Dataset Analysis/Biomathematics: Informatics

Contact Email