GaLNet: Weakly-Supervised Learning for Evidence-Based Tumor Grading and Localization in MR Imaging

Learning grading models directly from magnetic resonance imaging (MRI) without segmentation is more challenging. Existing deep convolutional neural network-based tumor grading algorithms rely on pixel-level annotation and lack of interpretability for clinical applications. This paper proposes a tumor Grading and Localization Network, or GaLNet, for providing evidence-based tumor grading from original magnetic resonance images (MRI) without tumor segmentation using a weakly-supervised approach. By employing malignancy attention blocks, GaLNet learns multi-scale malignancy-aware features with both strong semantics and fine spatial information. By adapting GaLNet trained with image-level tumor grading labels, the network jointly localizes malignant regions to provide supporting evidence of why and what GaLNet predicts. GaLNet achieves an AUC of 0.86 and an accuracy of 87% on testing dataset, which outperforms the tradition ResNet (0.83, 80%) and SENet (0.84, 86%) trained with images of segmented tumor.