MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel
  1. Home
  2. Meeting Program
  3. All
  4. Radiomics-Dosiomics Approach Improves Prediction of Radiation Pneumonitis Compared to DVH Data in Lung Cancer Patients

Radiomics-Dosiomics Approach Improves Prediction of Radiation Pneumonitis Compared to DVH Data in Lung Cancer Patients

N Chopra1*, T Dou2, R Mak3, G Sharp4, E Sajo5, (1) Massachusetts General Hospital, Boston, MA, (2) Texas Center for Proton Therapy, Irving, TX, (3) Brigham and Women's Hospital and Dana-Farber Cancer Center, Boston, ,(4) Massachusetts General Hospital, Boston, MA, (5) Univ Massachusetts Lowell, Lowell, MA

Presentations

(Tuesday, 7/14/2020) 1:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Room: Track 2

Purpose: To compare DVH-based model against the prognostic power of radiomics (CT) and dosiomics (TPS Dose Image) features using peritumoral dose distributions to predict radiation pneumonitis (RP) in a large cohort of 701 locally advanced non-small cell lung cancer patients.


Methods: CT and TPS thoracic dose data was used to generate pulmonary regions of interest (ROI) corresponding to two specific dose-levels (V20 and V5 ROIs) by subtracting the tumor volume from bilateral lung parenchyma receiving = 20 Gy or = 5 Gy respectively on the registered CT. Radiomics and dosiomics features were extracted for the V20 and V5 bilateral pulmonary ROIs from CT and 3D dose distribution using the open-source PyRadiomics toolkit. Only patients that completed a curative treatment course were included (median dose = 62 Gy). The endpoint was CTCAE grade>=2 RP (n=73). Feature selection was performed for each data set using the Minimum Redundancy Maximum Relevance (mRMR) algorithm and classification performance was evaluated for highly predictive features using the AUC statistic and compared to the various composite models including a clinical model as well as a dose-volume histogram based model. Clinical predictors considered in our study include age, gender, smoking, pack years, treatment modality, performance status. A random forest classifier was chosen to evaluate the multivariate model performance.


Results: While the dosiomics model performance was comparable to the DVH model, we find that a radiomics+dosiomics model with features extracted from the V20 ROI (AUC = 0.713) & combined V20+V5 radiomics+dosiomics model (AUC= 0.708) outperform the DVH only model (AUC = 0.63).


Conclusion: Pre-treatment evaluation of normal tissue regions using a radiomics-dosiomics approach shows promise in improving prediction of Grade>=2 RP as compared to conventional DVH metrics, thereby showcasing the potential of leveraging traditional machine-learning techniques in quantifying the risk of developing RP2 in lung cancer patients.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email

Additional Links