Encrypted login | home

Program Information

Multi-Modality Radiomics Model for Predicting Distant Failure in Lung SBRT

Z Zhou

Z Zhou*, M Folkert , P Iyengar , K Westover , H Choy , R Timmerman , S Jiang , J Wang , The University of Texas Southwestern Medical Center, Dallas, TX


MO-RAM-GePD-J(B)-3 (Monday, July 31, 2017) 9:30 AM - 10:00 AM Room: Joint Imaging-Therapy ePoster Lounge - B

Purpose: A new multi-modality radiomics model was proposed to predict distant failure in early stage non-small cell lung cancer (NSCLC) patients treated with stereotactic body radiation therapy (SBRT).

Methods: During constructing a predictive model, most radiomics-based approaches combine the features extracted from different modalities, such as PET and CT. However, as the imaging principles for different modalities are always different, the features extracted from different modalities also have different clinical/physiological indications. Therefore, simply combining all features together into a single predictive model may not be ideal. As such, a multi-modality radiomics model was proposed in this work. In the proposed multi-modality radiomics model, individual predictive model was first constructed for each modality. The final output was obtained by combining the probability output with the corresponding weight from all the modalities. During the model training, both sensitivity and specificity are taken as the objective functions simultaneously to reduce the effect of data imbalance. The new model is used to build a model for distant failure prediction in lung SBRT using 52 patients treated at our institute. Two image modalities including PET and CT, and clinical parameters were utilized as input for the model. The modelling procedure consists of three steps: (1) Extracting the features from the segmented tumors in PET and CT images; (2) Selecting the most reliable features for each modality from extracted features; (3) Training model and weight based on multi-objective.

Results: Mean and standard deviation of AUC, sensitivity, and specificity for multi-modality model from ten times running results are 0.83±0.01, 0.80±0.01, 0.89±0.01, respectively. The AUC, sensitivity, and specificity for a model by simply combining all the features are 0.80±0.01, 0.75±0.01, 0.87±0.01, respectively.

Conclusion: We developed a multi-modality radiomics model for predicting distant failure in NSCLC patients treated with SBRT, with improved performance compared to the current method.

Contact Email: