Automated Brain Tumor Segmentation on Multi-MR Sequences to Determine the Most Efficient Sequence using a Deep Learning Method

Abstract

Brain tumor segmentation is an important step in the diagnosis and treatment planning of cancer patients. The procedure of manual brain tumor segmentation suffers from a long processing time. In this light, automatic brain tumor segmentation is highly appealing in the clinical routine. This study sets out to segment the tumors from brain MR images and to investigate the effectiveness/usefulness of the different MRI sequences for this purpose. Here, the MR images from the BRATS challenge were utilized. 310 patients with four different MRI sequences, including T1, T1ce, T2, and FLAIR were employed to train a ResNet deep CNN. Four separate models were trained with each of the input MR sequences to identify the best sequence for brain tumor segmentation. To assess the performance of these models, 60 patients (external dataset) were quantitatively evaluated. The quantitative results indicated that the FLAIR sequence is more reliable for automatic brain tumor segmentation than other sequences with an accuracy of 0.77±0.10 in terms of Dice compared to Dice indices of 0.73±0.12, 0.73±0.15, and 0.62±0.17 obtained from T1, T2, and T1ce sequences, respectively. Based on the results of this study, FLAIR is a more reliable sequence than other sequences for brain tumor segmentation.

Keywords

• segmentation
• mri
• deep learning