Automatic segmentation of brain tumours in 3D MRI : an adaptation of 3D U-Net++

Abstract

In recent times, the capabilities and availability of deep learning techniques for image segmentation have increased considerably. U-Net is an example of a popular deep learning model designed specifically for biomedical image segmentation. U-Net++ combines the parameter efficiency of U-Net with the benefits of dense networks; namely the usage of more refined global and local feature contexts. We propose a model which adapts the U-Net++ architecture for brain tumour segmentation, with some of the modifications to the architecture resulting in an even smaller model. The proposed approach obtained Dice Coefficient scores of 0.7192, 0.8712, and 0.7817 for the Enhancing Tumour, Whole Tumour and Tumour Core classes when compared to the expert ground truths of the BraTS 2019 Challenge validation dataset. Whilst these results are surpassed by some published works discussed in Section 4.3.2, the results obtained are still fairly positive. The proposed approach differs from the standard U-Net++ model in a number of ways. Firstly, the optimisation function used in this experiment is a multiclass Dice Coefficient loss rather than the binary cross entropy. The number of convolutional blocks was halved to produce a much smaller model. U-Net++ also produces full-resolution secondary segmentation maps, which the proposed model combines using element-wise additions rather than averaging as in the original work. The model also makes use of 3 × 3 × 3 segmentation kernels instead of 1 × 1 × 1 as some initial tests showed that this configuration produced slightly better results. The experiments performed throughout the project’s development demonstrate the importance of using data augmentation and post-processing techniques, both of which substantially improved the model predictions. Furthermore, additional experiments were carried out including a change in optimisation function to the original binary cross entropy function used for U-Net++; using the standard number of convolutional blocks, and the use of dropout regularisation. These experiments were selected on the basis of them being either standard experiments (dropout) or properties of the original model (binary cross entropy, convolutional blocks). Nonetheless, this latter set of experiments did not merit sufficient improvement to be included in the proposed model architecture. Thus, this dissertation presents a novel adaptation of the U-Net++ architecture, which is lightweight and produces good quality segmentation results. A variant of the model was also submitted to the Organisation for Human Brain Mapping (OHBM) for a poster presentation, and to the IEEE MELECON 2020 Conference as a conference paper under the ’Advances in Medical Informatics for Healthcare Applications’ track. Both submissions were accepted by the respective bodies and presented using electronic methods due to the ongoing global pandemic situation.