Gradual Self‐Training via Confidence and Volume Based Domain Adaptation for Multi Dataset Deep Learning‐Based Brain Metastases Detection Using Nonlocal Networks on MRI Images

Background Research suggests that treatment of multiple brain metastases (BMs) with stereotactic radiosurgery shows improvement when metastases are detected early, providing a case for BM detection capabilities on small lesions. Purpose To demonstrate automatic detection of BM on three MRI datasets using a deep learning‐based approach. To improve the performance of the network is iteratively co‐trained with datasets from different domains. A systematic approach is proposed to prevent catastrophic forgetting during co‐training. Study Type Retrospective. Population A total of 156 patients (105 ground truth and 51 pseudo labels) with 1502 BM (BrainMetShare); 121 patients with 722 BM (local); 400 patients with 447 primary gliomas (BrATS). Training/pseudo labels/validation data were distributed 84/51/21 (BrainMetShare). Training/validation data were split: 121/23 (local) and 375/25 (BrATS). Field Strength/Sequence A 5 T and 3 T/T1 spin‐echo postcontrast (T1‐gradient echo) (BrainMetShare), 3 T/T1 magnetization prepared rapid acquisition gradient echo postcontrast (T1‐MPRAGE) (local), 0.5 T, 1 T, and 1.16 T/T1‐weighted‐fluid‐attenuated inversion recovery (T1‐FLAIR) (BrATS). Assessment The ground truth was manually segmented by two (BrainMetShare) and four (BrATS) radiologists and manually annotated by one (local) radiologist. Confidence and volume based domain adaptation (CAVEAT) method of co‐training the three datasets on a 3D nonlocal convolutional neural network (CNN) architecture was implemented to detect BM. Statistical Tests The performance was evaluated using sensitivity and false positive rates per patient (FP/patient) and free receiver operating characteristic (FROC) analysis at seven predefined (1/8, 1/4, 1/2, 1, 2, 4, and 8) FPs per scan. Results The sensitivity and FP/patient from a held‐out set registered 0.811 at 2.952 FP/patient (BrainMetShare), 0.74 at 3.130 (local), and 0.723 at 2.240 (BrATS) using the CAVEAT approach with lesions as small as 1 mm being detected. Data Conclusion Improved sensitivities at lower FP can be achieved by co‐training datasets via the CAVEAT paradigm to address the problem of data sparsity. Level of Evidence 3 Technical Efficacy Stage 2