Purpose: Knowledge of tumor growth behavior is key to clinical decision-making in the management of vestibular schwannoma (VS) patients. In clinical practice, a series of images acquired at a fixed interval are used to determine VS tumor growth, which may lead to delayed treatment for fast-growing tumors or over-imaging for slow-growing tumors. In this work, we aim to build an attention guided convolutional neural network to predict VS tumor growth with the initial diagnosis imaging to aid in personalized management of VS.
Methods: Contrast-enhanced T1-weighted MR images of 103 VS patients were used in this study. Each patient has two MRI scans with median interval of 7.2 months. VS on these images were segmented by a neurotologist. Sixty-two out of 103 patients experienced VS growth based on the volumetric measurements on two consecutive scans. Our attention guided model used an UNet segmentation module and an AlexNet-style classification module to exploit the tumor growth related structures for VS growth prediction. The segmentation module is used to identify tumor and its surrounding area relevant to tumor growth prediction and generate attention maps from different layers, while the classification module takes the MR image and attention maps as input and predicts tumor growth. These two modules are trained simultaneously through optimizing the combination of dice coefficient loss from the segmentation part and binary cross-entropy loss from the prediction part.
Results: Eighty patients’ data were used for model training, 10 for validation, and 13 for testing. The growth prediction accuracy and AUC of the proposed model on testing data are 0.85 and 0.82, which are higher than those from a separate classification network without the attention mechanism.
Conclusion: We developed an attention guided network for VS tumor growth prediction using the initial diagnosis MR image. Comparison with separate modules demonstrated its efficiency and effectiveness.
IM/TH- Image Analysis (Single Modality or Multi-Modality): Computer-aided decision support systems (detection, diagnosis, risk prediction, staging, treatment response assessment/monitoring, prognosis prediction)