Room: Track 4
Purpose: metastases are one of the most common neurologic complications of cancers. Stereotactic radiosurgery (SRS) is a well-established treatment for brain metastases, which administers a highly conformal high-dose irradiation of each metastasis and hence requires accurate detection and delineation of metastases. However, manually detecting all the brain metastases can be labor-intensive and time-consuming, given their small sizes and multiplicity. We propose to employ 3D Mask R-CNN to automatically detect brain metastases on magnetic resonance (MR) images for SRS treatment planning.
Methods: the training stage, coarse feature maps were extracted from 3D MRI patches using a pretrained ResNet. A region proposal network was then employed to predict the locations and sizes of candidate volumes of interest (VOIs) from these feature maps. Suspect metastases within candidate VOIs were then segmented using a uniformed fully convolution network. Segmentation loss, classification loss, as well as VOI location and size regression loss were used to supervise our networks. After training, to detect brain metastases for a new patient, the extracted patches of MRI were fed into our trained model to predict candidate VOIs and metastases probability maps within the VOIs, and then re-transformed and resampled them to the original image resolution. A consolidation was performed on these predictions via weighted cluster scoring to determine the final VOIs with proper sizes.
Results: have tested our method on 20 patients’ brain contrast T1-weighted MRI images using five-fold cross-validation, and achieved 86.5%±3.2% sensitivity and 89.7%±4.8% specificity. For each patient, it took our trained model a few seconds to detect the brain metastases on the 3D MRI images.
Conclusion: preliminary study has demonstrated the efficacy and clinical feasibility of our auto-detection method, implying its potential to significantly improve the efficiency of SRS treatment planning and hence ultimately improve the clinical outcome.