Click here to


Are you sure ?

Yes, do it No, cancel

Treatment Response Prediction Using Texture Features From Longitudinal Diffusion MRI in Sarcoma Patients

Y Gao1*, C Gu1,2 , J Kim1 , M Cao1 , J Fu1, A Kalbasi1 , D Ruan1 , J Lewis1 , D Low1 , P Hu1 , Y Yang1 , (1) UCLA, Los Angeles, CA, (2) Xi'an Jiaotong University, Xi'an, Shaanxi,


(Wednesday, 8/1/2018) 7:30 AM - 9:30 AM

Room: Karl Dean Ballroom C

Purpose: To predict the treatment response using features extracted from longitudinal diffusion MRI.

Methods: Twenty-four soft-tissue sarcoma patients treated with five-fraction radiotherapy treatment were recruited. Diffusion images were acquired three times throughout the treatments (T1-before the first treatment, T2-in the middle, and T3-after the last treatment) using a 0.35T MRI-guided radiotherapy system. A necrosis score ranging from 0% to 100% was obtained from the post-radiotherapy resection as an immediate surrogate of the treatment response. Patients were divided into N_low and N_high group based on the necrosis score(<50% v.s. ≥50%). Thirty-five features were extracted from the tumor ADC maps at each timepoint. Temporal difference and ratio between different timepoints were included (total of 315 features). Minimum redundancy maximum relevance(MRMR) was used for feature selection. Logistic regression(LR), support vector machine(SVM) and adaptive boosting(AdaBoost) were implemented to predict treatment response. Six-fold cross-validation was repeated 50 times with different data splitting during both feature selection and model training stages to improve and estimate the model robustness. Classifications using features from all three timepoints (T1-3), from single timepoint (T1, T2, T3) and from initial and end timepoints (T1+T3) were performed.

Results: Using features from all three timepoints provided the best performance. Among the three models, SVM outperformed the other two for all statistics (AUC=0.86±0.04, sensitivity=0.86±0.07, specificity=0.87+0.06, accuracy=0.86±0.04.). AUCs for LR and AdaBoost were 0.74±0.06 and 0.80±0.05, respectively.The highest AUC using features from T1+T3 was achieved with AdaBoost (AUC=0.78±0.06). Prediction performance drastically reduced when only a single timepoint was used, which might be an indication that longitudinal data is essential for treatment response prediction. The best AUC was all achieved with SVM: AUC_T1=0.65±0.07, AUC_T2=0.67±0.06, and AUC_T3=0.65±0.05.

Conclusion: Longitudinal diffusion MRI was used for the prediction of necrosis score on sarcoma patients. The SVM model with features from all timepoints provided the best performance.


Not Applicable / None Entered.


IM- MRI : Quantitative Imaging/Analysis

Contact Email