Program Information
Automatic Segmentation Refined, Multiple Slice-Wise Voting Based Classification of Tumors From MRI
H Veeraraghavan*, Y Lakhman , D Feier , H Vargas , H Hricak , J Deasy , E Sala , Memorial Sloan Kettering Cancer Center, New York, NY
Presentations
SU-E-J-255 (Sunday, July 12, 2015) 3:00 PM - 6:00 PM Room: Exhibit Hall
Purpose: Development of technical methods to optimally classify tumors via magnetic resonance imaging.
Methods: The methods were developed and tested on unusual appearing leiomyomas (ULM) from leiomyosarcomas (LMS). We developed a fully automated method for distinguishing between ULM and LMS from T2-w MR images. Data consisted of 39 patients with histologically proven ULM(=22) and LMS(=17) who underwent preoperative ≥1.5 MRI. 13 MR images were obtained at our institution and rest from elsewhere. Our method consists of several steps. First, all the images were histogram matched, following which the manually segmented tumors were refined through automatic volumetric image segmentation. Next, several image features consisting of 5 Haralick textures, 4 Gabor edges at (0°,45°,90°,135°) and bandwidth 1.414 and Haralick textures computed on each Gabor image resulting in 25 different features was computed from inside the segmented tumors. The features were pre-weighted by their relevance determined using a paired t-test and trained using a random forest classifier. We improved the generalization performance of the classifier by boosting the number of samples by using slice-wise computed textures resulting in a total of 604 samples with ULM (n=351) and LMS (n=253). In the slice-wise method, classifier is built that applies to individual slices; the slice results are then combined using voting to determine an overall per-patient classification.
Results: We evaluated the performance of the classifier for volumetric texture-based classification with Leave One Out Cross Validation and slice-wise textures-based classification with K=5 fold classification. The volumetric method achieved a classification accuracy of 72% in comparison to 87% using slice-wise texture-based classification. Gabor textures were more relevant than intensity-based textures for distinguishing the ULM from LMS.
Conclusion: We have demonstrated that a classifier method using automatic refinement of human drawn contours, along with voting on slice-based classification, results in an effective tumor image classifier.
Contact Email: