Program Information
Multi-Classifier Radiomics Model for Predicting Distant Failure in Cervical Cancer Using PET Image Features
Z Zhou*, G Maquilan , K Thomas , M Folkert , K Albuquerque , J Wang , UT Southwestern Medical Center, Dallas, TX
Presentations
TU-H-CAMPUS-JT-2 (Tuesday, August 1, 2017) 4:30 PM - 5:30 PM Room: Joint Imaging-Therapy ePoster Theater
Purpose: To develop a multi-classifier radiomics model and a reliable classifier fusion strategy for distant failure prediction in cervical cancer after radiotherapy.
Methods: Current radiomics models typically utilize a single classifier to construct a predictive model. However, if one classifier is considered as one “expert”, the performance of radiomics models can be potentially improved by combining decisions from multiple “experts”. In this work, a multi-classifier radiomics model was constructed by combining multiple classifiers. When combining the probability output from each classifier, current methods don’t consider the effect of output labels from other classifiers, which may lead to an unreliable probability output. Therefore, a new reliable classifier fusion strategy based on the evidential reasoning rule was proposed, where both the reliability and weight were combined to obtain the final probability output. Furthermore, when training the model, multiple objectives including both sensitivity and specificity were taken as the objective functions to overcome the imbalance problem of the training dataset. The proposed multi-classifier model is used to predict distant failure in 75 cervical cancer patients using imaging features extracted from PET. Six classifiers were used for constructing the predictive model, while an iterative multi-objective immune algorithm (IMIA) was used for solving the multi-objective optimization.
Results: Mean and standard deviation of AUC, sensitivity, and specificity for the proposed model from ten times running results are 0.83±0.02, 0.79±0.00, 0.84±0.03, respectively. The average AUC, sensitivity and specificity of six individual classifiers are 0.74, 0.75 and 0.74 respectively.
Conclusion: A new multi-classifier radiomics model for predicting distant failure in cervical cancer was developed. By combining the multiple classifier probability outputs, the proposed strategy outperformed any single classifier based model.
Contact Email: