Fully Convolutional Neural Network for Improved Brain Segmentation

Brain segmentation is key to evaluating brain structure for disease diagnosis and treatment. Much research has studied the segmentation of brain images. However, prior research has paid little attention to separating actual brain pixels from those related to the background of brain images. Failure to perform such a separation may (a) distort brain segmentation models and (b) introduce overhead to the modeling performance. In this paper, we improve the performance of brain segmentation using a 3D, fully convolutional neural network (CNN) model. We propose (i) a multi-instance loss method to separate actual brain pixels from background and (ii) Gabor filter banks with K-means clustering to provide informative segmentation features. We provide deeper analysis and discussion of evaluation results and the state-of-the-art models. Evaluated on infant and adult datasets, our model achieves dice coefficients of 87.4-94.1%, an improvement of up to 11% to the results of five state-of-the-art models. Unlike previous studies, we consult experts in medical imaging to evaluate our segmentation results. Feedback from experts reveals that our results are fairly close to the manual reference. Moreover, we observe that our model is 1.2x-2.6x faster than prior models. We conclude that our model is more accurate and efficient in practice for segmenting both infant and adult brain images.

Automated summary

This paper proposes an improved method for brain segmentation using a 3D convolutional neural network model. The method includes multi-instance loss and Gabor filter banks. Evaluation results show that the model is more accurate and efficient in segmenting brain images than existing models.