D.S. Davis¹*, (1) University of South Carolina, Columbia, SC

• 
  D Davis

PO-BPC-Virtual-9 (Saturday, 4/17/2021)   [Eastern Time (GMT-4)]

Purpose: To survey deep learning and image segmentation techniques that automatically delineate multiple brain metastases within 3D MRI images.

Methods: Six different techniques using deep learning and automatic image segmentation of multiple brain metastases detected in 3D MRI images were found in the literature. The following results were collected from each article: depth of deep learning algorithm, segmentation algorithm, DNN activation function, size of tumors, sensitivity and dice similarity performance scores, and final output produced.

Results: Deep learning neural networks, DNN, ranged in depth from 11 layers to 101 layers. The reported sensitivity and false positive rates ranged from 0.77 to 0.98 and the reported dice similarity coefficients ranged from 0.70 to 0.93. Brain metastases smaller than 0.40ml were reported challenging to detect and delineate automatically with most. One solution split the segmentation process into multiple sizes and then applied classification algorithm, En-DeepMedic. Additionally, each study independently built training data from retrospective patient chart review, as there is no benchmark datasets available for this highly specialized radiation therapy field. When results were reported, the range of sizes of the brain metastases were different as they were unique to patients sampled.

Conclusion: In this research, data augmentation plays an important role in the training process for treatment site. No one automatic segmentation technique had been adopted for SRS. The solution to this problem continues to evolve with technological advances in computer vision and multiple object detection. The most recent algorithm surveyed, Mask R-CNN, has not been thoroughly researched as a viable solution.

Keywords

Image Analysis, Segmentation, Stereotactic Radiosurgery

Taxonomy

IM- MRI : Machine learning, computer vision

Contact Email