Abstract
The purpose of this study was to develop and verify a convolutional neural network (CNN)-based deep-learning algorithm to identify tumor progression versus response by adding amide proton transfer-weighted (APTw) MRI data to structural MR images as the proposed model input. 145 scans with 2175 MR instances from 98 patients with malignant glioma (acquired between April 2010 and February 2018) were re-analyzed. An end-to-end classification framework based on a ResNet backbone was developed. The architecture includes a learnable subtraction layer and a hierarchical classification paradigm, and synthesizes information over multiple MR slices using a long short-term memory. Areas under the receiver-operating-characteristic curves (AUCs) were used to assess the impact of adding APTw MRI to structural MRI (T(1)w, T(2)w, FLAIR, and GdT(1)w) on classification of tumor response vs. progression, both on the slice- and scan-level. With both APTw and structural MRI data, adding a learnable subtraction layer and a hierarchical classification paradigm to the backbone ResNet model improved the slice-level classification performance from an AUC of 0.85 to 0.90. Adding APTw data to structural MR images as input to our proposed CNN classification framework led to an increase in AUCs from 0.88 to 0.90 for the slice-level classification (P < 0.001), and from 0.85 to 0.90 for the scan-level classification (P < 0.05). Generated saliency maps highlighted the vast majority of lesions. Complementing structural MRI sequences with protein-based APTw MRI enhanced CNN-based classification of recurrent glioma at the slice and scan levels. Addition of APTw MRI to structural MRI sequences enhanced CNN-based classification of recurrent glioma at the slice and scan levels.