Multi-modal MRI segmentation of sarcoma tumors using convolutional neural networks

Abstract

Small animal imaging is essential in building a bridge from basic science to the clinic by providing the confidence necessary to move new cancer therapies to patients. However, there is considerable variability in preclinical imaging, including tumor volume estimations based on tumor segmentation procedures which can be clearly user-biased. Our group is engaged in developing quantitative imaging methods which will be applied in the preclinical arm of a co-clinical trial studying synergy between anti-PD-1 treatment and radiotherapy using a genetically engineered mouse model of soft tissue sarcoma. This study focuses on a convolutional neural network (CNN)-based method for automatic tumor segmentation based on multimodal MRI images, i.e. T1 weighted, T2 weighted and T1 weighted with contrast agent. Our images were acquired on a 7.0 T Bruker Biospec small animal MRI scanner. Preliminary results show that our U-net structure and 3D patch-wise approach using both Dice and cross entropy loss functions delivers strong segmentation results. We have also compared single performance using only T2 weighted versus multimodal MR images for CNN segmentation. Our results showthat Dice similarity coefficient were higher when using multimodal versus single T2 weighted data (0.84 ± 0.05 and 0.81 ± 0.03). In conclusion, we successfully established a segmentation method for preclinical MR sarcoma data based on deep learning. This approach has the advantage of reducing user bias in tumor segmentation and improving the accuracy and precision of tumor volume estimations for co-clinical cancer trials.