SU-D-BRA-03: What Image Features Are Useful for Tumor Segmentation in Multi-Modal Images?

Authors

  • Hu Y,

    1. Memorial Sloan-Kettering Cancer Center, New York, NY
    2. City College of New York, New York, NY
    3. Memorial Sloan-Kettering Cancer Center, New York, NY
    Search for more papers by this author
  • Grossberg M,

    1. Memorial Sloan-Kettering Cancer Center, New York, NY
    2. City College of New York, New York, NY
    3. Memorial Sloan-Kettering Cancer Center, New York, NY
    Search for more papers by this author
  • Mageras G

    1. Memorial Sloan-Kettering Cancer Center, New York, NY
    2. City College of New York, New York, NY
    3. Memorial Sloan-Kettering Cancer Center, New York, NY
    Search for more papers by this author

Abstract

Purpose:

We hypothesize that derived image features will improve accuracy of segmenting gross tumor volume (GTV) using (semi-)automatic methods when compared with raw intensities. Features from multi-modal images, however, require more CPU time and memory, for extraction, training and classification. In this study, we examine what image features are useful for segmenting GTV in multi-modal images.

Methods:

Eight cases of high grade brain tumor from MICCAI 2013 for brain tumor segmentation competition were examined. In each case, images of co-registered four MRI modalities: Flair, T1, T1 contrast and T2, as well as the ground truth are provided in the data set. Three cases were used for probing important features and the other five cases were for evaluation of segmentations with the selected features. Besides original image intensity, additional 7 candidate features were extracted: minimum, maximum, gradient, entropy, spot, ripple and edge, resulting in a 32-D (4 × 8) feature vector for each voxel. For selecting important features, a random forest with 100 decision trees was trained using 32-D features from the 3 probing cases. The top ranked features reported by the random forest were selected and used in a Conditional Random Fields (CRF) based semi-automatic method we previously developed to segment the 5 testing cases. Dice coefficent metric to the ground truth are reported and compared to intensity-only segmentation using the same segmentation method.

Results:

The feature selection random forest ranked intensity, minimum, maximum and entropy the most important features across all 4 MRI modalities. With the selected features (16-D), the CRF segmentation of 5 test cases yielded Dice coefficent 0.847±0.114 (mean±1SD), significantly better (p =0.016) than the Dice coefficient 0.797±0.135 of intensity-only (4-D) CRF segmentation.

Conclusion:

Multi-modal image features selected by a machine learning method can improve the accuracy of brain tumor segmentation significantly when compared with intensity-only features.

Research supported by Varian Medical System

Ancillary