Machine Learning Based Detection of Brain Cancer Metastases in MR Images for Stereotactic Radiosurgery Using Single-Shot Detectors

Abstract

Identification of brain cancer metastases in MR images for stereotactic radiation treatment is presently performed manually by a radiologist/radiation oncologist and can be time-consuming. The purpose of this study was to investigate machine learning based single-shot detectors (SSDs) for automatic detection of brain metastases. Post-contrast 3D T1-weighted gradient echo images (FOV=24 cm, TR/TE=6.947/2.428 ms, voxel size=0.94×0.94×1.00 mm) from 146 patients who underwent noninvasive stereotactic radiosurgery at our institution were included in the study. The manual identification and contouring of the brain metastases by neuroradiologists and treating radiation oncologists were used as the ground truth. The patients were partitioned into 3 cohorts of 99/25/22 for training/validation/testing purposes with a total of 438/117/93 brain metastases, respectively. The size of the brain metastases ranged from 1.88 to 50.63 mm and was distributed similarly in the 3 cohorts. We constructed a multibox SSD with 16 convolutional layers, of which 6 layers of different resolution scales (128×128, 64×64, 32×32, 16×16, 8×8 and 6×6) were used for prediction. We used the SSD loss function with a weighted sum between the localization loss (Smooth L1) and the classification loss (Softmax). Inputs to the network were the axial slices of the MR images, and the outputs were the bounding box coordinates and the associated confidence of the classifier. An Adam optimizer was used to train the network and the initial learning rate was set to 2×10-4. Random affine transformation augmentations, decayed learning rates and an early stopping scheme were applied during the training procedure. The detection performance was analyzed on 11 different confidence thresholds ranging from 99% to 0%. For each threshold, the predicted volumes of interest were reconstructed using the output bounding box coordinates and were compared with the ground truth. True positives (TPs) were defined as predictions having at least one voxel overlap with the ground truth bounding box; otherwise, they were considered false positives (FPs). An average sensitivity of 73.5% and FP/TP ratio of 15.8 were achieved across all sizes of brain metastases and all confidence thresholds. For metastases ≥6 mm, the average sensitivity was 94.5%, and the FP/TP ratio was 7.9. Detection performance for 30%, 60%, and 90% confidence thresholds are listed in Table 1. The SSD was able to detect almost all brain metastases ≥6 mm with high confidence and specificity, and nearly half the metastases <6mm with low specificity. The performance of the model is expected to further improve through ablation studies, hyperparameter searches, and inclusion of additional curated patient data.Abstract 1002; Table 1Detection performance of the SSD for 3 confidence thresholds.Confidence threshold<6 mm≥6 mmSensitivityFP/TP ratioSensitivityFP/TP ratio90%33.3%1.7190.2%0.9160%42.9%3.0696.1%1.8230%54.8%5.8796.1%3.69 Open table in a new tab