MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

An Empirical Comparison of Weka Classifiers for Outcome Prediction Using An Imaging Habitats Definition and Feature Extraction Method On MRI

Q Han1*, R Palm2, K Latifi2, E Moros2, A Naghavi2, G Zhang2, (1) University of South Florida, Tampa, FL, (2) H. Lee Moffitt Cancer Center, Tampa, FL

Presentations

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

Room: AAPM ePoster Library

Purpose:
We propose an Imaging Habitats Definition and Feature Extraction method on MRI and a practical method for medical researchers to build classification models from a small imbalanced dataset using Weka. We hypothesize that utilizing pretreatment MRI can predict outcome in Soft Tissue Sarcoma (STS) treated with neoadjuvant radiotherapy (RT).

Methods:
A total of 97 STS patients were included, with 68 and 29 in the training and testing cohorts, respectively. Gross-tumor volumes (GTV) were manually segmented on pre-RT T1-post contrast and T2 STIR sequences. The habitats method dichotomized the MRI images by signal intensity to form 4 imaging habitats: T1 high-T2 high, T1 high-T2 low, T1 low-T2 high and T1 low-T2 low; 154 habitat features were extracted, and 11 clinical features recorded. Multi-step dimensionality reduction and Weka machine learning classifiers were utilized to identify habitat features predictive of pathological necrosis rates (PNR) >90% or >95% at the time of surgery. Standard Weka classifiers and their combinations were tested. Performance metrics (Accuracy, AUROC, AUPRC, F-Measure, MCC) were computed.

Results:
Support Vector Machine with Stochastic Gradient Descent gave the best performance. A combination model of habitat and clinical features provided better prediction. An 8-feature model, 4 habitat and 4 clinical features, predicted for PNR>90% (training: 72%, ROC 0.746; testing: 65%, ROC: 0.713). Evaluating patients treated with neoadjuvant RT alone (n= 84), an 8-feature model with 5 habitat and 3 clinical features had the highest prediction for PNR>95% (training: 78%, ROC 0.721; testing: 72%, ROC 0.600).

Conclusion: approach we employed proved useful when using a relatively small and imbalanced sample to build classification models with clinically useful prediction power. Radiomic features suggestive of habitat diversity within a tumor appear to be associated with RT outcome.

Keywords

Image Analysis, MRI, Modeling

Taxonomy

IM- MRI : Machine learning, computer vision

Contact Email