Room: AAPM ePoster Library
Purpose: To evaluate the use of a 3D U-Net convolutional neural network (CNN) framework for automatic detection and segmentation of brain metastases in MR images.
Methods: Eighty-nine MRI datasets from patients with at least one brain metastasis and no previous irradiation from our institution were retrospectively analyzed. Patients were imaged using an MRI protocol that included a T1-weighted (T1w) post-gadolinium contrast agent image and a T2-weighted (T2w) image prior to receiving GammaKnife radiosurgery. Brain metastases were either contoured or reviewed on MRI by an expert neurosurgeon. The complete workflow employed was as follows: 1.Bias field correction of MRI images. 2.-Resampling of T2w image to T1w image resolution and registration of T2w to T1w image. 3.-Random split of the dataset on training set (n=72) and validation set (n=17). 4.-Training of the CNN using both sets of images (as different channels) and a binary mask of the brain metastases. The model was trained with a limit of 500 epochs using a high-performance computing cluster. 5.-Running the trained model on the validation set to obtain predicted brain metastases segmentation.
Results: Predicted metastases were analyzed based on size to establish a threshold for high sensitivity. Using a lesion volume threshold of 65 mm³ (5 mm equivalent sphere diameter) the sensitivity was 90.3% (28/31) and the false positive ratio per predicted lesion was 15.2% (5/33). Decreasing the lesion volume threshold to 14 mm³ (3 mm equivalent sphere diameter), the sensitivity dropped to 54.7% (29/53), and the false positive ratio per predicted lesion increased to 35.6% (16/45) clearly indicating that the trained model fails to detect lesions of smaller size.
Conclusion: Deep learning holds promise for automatic detection and segmentation of brain metastases. Our results indicate that high detection rates can be obtained for metastases that are at least 5 mm in size using our method.