Encrypted login | home

Program Information

Non-Small-Cell Lung Cancer (NSCLC) Tumor Segmentation On MRI Images with Multiple Sequences Using Learning Combined with Semi-Automatic Segmentation Methods

no image available
Y Hu

Y Hu1*, H Veeraraghavan1 , J Jiang1 , N Tyagi1 , G Mageras1 , (1) Memorial Sloan Kettering Cancer Center, New York, NY


SU-K-FS4-2 (Sunday, July 30, 2017) 4:00 PM - 6:00 PM Room: Four Seasons 4

Purpose: The utility of MRI for lung tumor delineation for radiation treatment planning has become plausible with advances in MRI imaging techniques. We hypothesize that exploiting multiple MRI sequences will improve the performance of computer-assisted segmentation. We developed and evaluated two multi-modality semi-automatic methods combined with learning that simultaneously utilize multiple MRI sequences for NSCLC tumor segmentation.

Methods: The user applies brush strokes to indicate primary tumor and non-tumor tissues for guiding tumor segmentation of NSCLC cases. Image intensities from respiratory-triggered T2-weighted (T2w) and Inversion Recovery (IR) MRI sequences under the brush strokes are used to train patient-specific Support-Vector-Machines (SVMs) which were then used to classify all voxels in the image to produce voxel labels and label probability. We developed two different approaches for combining the SVM with semi-automatic segmentation, consisting of: (1) tumor segmentation confirmation wherein SVM-based tumor labels were combined with interactive grow-cut method (GrowCut+SVM), and (2) computation of SVM-estimated boundary potentials in Conditional Random Field (CRF) to drive a graph cut segmentation (SVMCRF). Evaluation used Dice coefficients between manual and predicted segmentations from 5 patient cases to compare the performance of the 2 methods.

Results: For the GrowCut+SVM method, Dice coefficients (Mean±SD) were 0.48±0.18 with T2w only, 0.70±0.13 with IR only, and 0.72±0.11 with T2w and IR together. For the SVMCRF method, coefficients were 0.73±0.12, 0.72±0.13, and 0.76±0.12, respectively. SVM models trained with both T2w and IR improved the accuracy in 4 of 5 cases for GrowCut+SVM and SVMCRF methods. Growcut+SVM on T2w performed significantly worse than other cases (p<0.05). In general, SVMCRF yielded higher Dice coefficients.

Conclusion: This preliminary study indicates that using SVM models with multiple MRI sequences may help segmentation performance compared to models with single sequence alone. In addition, among the two semi-automatic methods studied, the CRF-based method exhibited less leakage.

Funding Support, Disclosures, and Conflict of Interest: supported by Varian Medical System

Contact Email: