Our goal is to predict treatment outcome for radiation therapy using image features from dynamic contrast enhanced magnetic resonance imaging (DCE-MRI). We are using image data collected from 23 patients with cervical cancer who underwent radiation therapy.
18 patients with an outcome of local control of the tumor.
5 patients with an outcome of local recurrence.
2 DCE-MRI imaging studies performed for each patient, the first prior to treatment and the second 2-3 weeks into treatment. Each study consisted of a pre-contrast series and a post-contrast series.
Figure 1 : Pre-contrast
Figure 2 : Post-contrast
Tumor ROIs drawn manually.
12 studies required registration – these linear registrations were performed manually.
In order to reduce the partial volume effects, the outermost pixels along the perimeter of each region were removed from the ROIs.
Calculate tumor features (average intensity, etc.). Test feature subsets for class separability based on treatment outcome.
Reduce dimension of subsets to 2 dimensions by principal components analysis (PCA).
(Left) Initial alignment of Anatomical scan (grayscale image) and DCE scan (green outline) and (Right) registered alignment
Best performing classifier used pre-therapy and early therapy mean voxel intensities, both before contrast injection, with early therapy intensity distribution skewness, after contrast injection. The prediction accuracy using the support vector machine (SVM) and linear discriminant analysis (LDA) classifiers were 73.9%±13.6% and 68.5%±18.3%, respectively.
Treatment outcomes (Local Control (LC) or Local Recurrence (LR)) plotted after PCA projection of feature set
Plot showing linear classifier separation of treatment outcome classes using support vectors (encircled). Notice improperly classified points in blue near top of plot