Contrastive Learning with Dynamic Weighting and Jigsaw Augmentation for Brain Tumor Classification in MRI

Magnetic resonance imaging (MRI) is considered the most effective non-invasive technique to help diagnose brain tumor diseases. Recently, computer-aided diagnostic models based on deep convolutional neural networks have provided efficient assistance for brain disease detection from MRIs. However, the performance of these supervised models relies heavily on labeled data. And data annotation work is error-prone in practical clinic applications because it requires expert domain knowledge and significant time investment. Self-supervised models hold great promise in training unlabeled data except for the weaknesses such as an unfriendly computing budget and inferior performance. Thus, we propose a novel self-supervised framework for the unsupervised classification of brain MRIs. Firstly, we build an interleaved structure for contrastive learning, aiming to facilitate the backbone network to seek a better representation with a low computational cost. Then we apply a dynamic weighting mechanism to prevent collapsed solutions and pursuit a perfect trade-off between convergence speed and high performance. Moreover, we introduce a jigsaw puzzle solver as an additional augmentation tool, aiming to collect more valuable clues for correct classes by expanding the attention to local spatial information. Comprehensive experiments show the superiority of the proposed model through qualitative and quantitative evaluation.