Effect of anatomical contextual information on the performance of a convolutional neural network tasked with brain metastasis detection

Abstract Purpose : Modern management of brain metastases requires close surveillance with serial magnetic resonance imaging (MRI), creating interest in techniques for automated detection. The convolutional neural networks (CNN) is currently the dominant approach to automated metastasis detection, and multiple variations of the CNN have been investigated to improve performace. In this work, we looked beyond the impact of network architecture and assessed the impact on performance of providing anatomical contextual information to a CNN in the training period. Materials and Methods : The nn-Unet, a widely-adopted CNN, was selected for this study. The nn-Unet was trained on an institutional dataset consisting of 301 high-resolution, T1-weighted, contrast-enhanced MRIs and associated structure labels consisting of both target brain metastases (1,111 total) and several organs at risk (OARs) that were labeled in the process of radiosurgery planning. In the training period, the model was presented either with labeled metastases alone (BM) or with the complete structure set of metastases and OARs (BM+OAR). The test set contained 100 cases and 421 total metastases. Results : We found that by multiple standard performance metrics, the BM+OAR model outperformed the BM model. Median case-wise Dice Coefficient for BM was 0.75 (IQR 0.51-0.86) and for BM+OAR was 0.79 (IQR 0.58-0.86) (p<0.001). Median sensitivity for BM was 1.0 (IQR 0.71-1.0) and for BM+OAR was 1.0 (IQR 0.75-1.0) (p<0.01). Median positive predictive value for BM was 0.5 (IQR 0.25-0.74) and for BM+OAR was 0.77 (IQR 0.50-1.0) (p<0.001). Importantly, amongst lesions scored as "false-positives", 47/151 detected by BM+OAR could be retrospectively verified to be real tumors while only 22/231 detected by BM could be verified to be real (p<0.001). Conclusions : These findings suggest that providing anatomical contextual information can improve accuracy of CNN-based automated brain metastasis detection algorithms and opens new opportunites for research into optimization of automated detection approaches for brain metastases.