Radiomic Features Predict Rib Fracture after Stereotactic Body Radiotherapy for Early Stage Non-Small Cell Lung Cancer

Abstract

a 3D convolutional neural network can auto-segment the lens, optic nerves and chiasm on planning CT scans alone for Gamma Knife SRS with similar accuracy to manual contours. Materials/Methods: A dataset of 101 patients treated with Gamma Knife SRS at a single institution for pituitary tumors or skull base meningiomas were collected. High resolution CT and MRI images were fused and the optic nerves and chiasm were contoured by a neuroanatomic expert. These contours were then transferred to the CT image. The dataset was split into training, validation, and test sets and the data was trained using an attention-gated 3D Residual Network (AG-3DResNet) with a generalized dice loss function. The predicted structures were then compared to manual contours using the Dice Similarity Coefficient (DSC), 95% Hausdorff Distance (HD), Precision, Recall, and Relative Average Volume Difference (RAVD). Results: We compared the predicted contours from the model to the manually constructed contours for all 101 patients. Conclusion: Our AG-3DResNet deep convolutional network can provide accurate segmentation of the lens, optic nerves and chiasm on planning CT scans alone for Gamma Knife SRS. The Dice and 95% HD for our model outperformed the current published best performing model, AnatomyNet, for the Chiasm and Optic Nerves by nearly 10%.