Purpose: Stereotactic radiosurgery requires precise target delineation, which can be highly time-consuming especially in patients with multiple (>4) brain metastases (BMs). We have developed a CNN-based BMs segmentation platform (BM-Seg) for accurate and efficient BMs delineation, but it suffers from false-positives caused by varied levels of contrast in planning MRI acquisition. Built on BM-Seg, we developed false-positive reduction Siamese network (FPrSN) to improve BM-Seg performance.
Methods: The proposed FPrSN consists of a parallel feature extractor and a following classifier. The feature extractor utilized a pair of truncated pre-trained ResNet50 with shared weights. In the training stage, paired BM candidate images are used as the input into the feature extractor to 1) increase the training set size by a factor of N/2 and 2) identify the inter-class difference. Then the classifier will randomly take the output from one branch of the feature extractors to classify the corresponding input as a false-positive segmentation or true BM. In the testing stage, only one branch of the feature extractor will be utilized with the classifier to classify a single candidate input as false-positive segmentation or true-positive segmentation. The performance of FPrSN and FPrSN-integrated BM-Seg are evaluated.
Results: The performance of the FPrSN reaches accuracy of 0.94, area under the curve of 0.92, sensitivity of 0.88 and specificity of 0.97 in a testing dataset. The BM-Seg has initial sensitivity of 0.96 and specificity of 0.59 in BMs detection/segmentation, and get improved to 0.86 and 0.89 after integrating FPrSN, respectively. Due to the flexibility of FPrSN, sensitivity/specificity balance of the FPrSN-integrated BM-Seg can be further adjusted to satisfy clinic needs.
Conclusion: This developed FPrSN can effectively reduce the false-positive rate of the BMs segmentation candidates. Integrating FPrSN in the BM-Seg can improve the performance, and this integrated platform could be a beneficial tool for clinical implementation.
Funding Support, Disclosures, and Conflict of Interest: This work is supported by NIH R01 CA235723 and the seed grand from the Department of Radiation Oncology of the UT Southwestern Medical Center.
Segmentation, Stereotactic Radiosurgery, CAD
IM/TH- Image Segmentation Techniques: Clustering, classification and threshold