MR brain tissue segmentation is a significant problem in biomedical image processing. The goal is to segment images into three tissues, namely white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF). We use a LSTM method with multi-modality and adjacency constraint for brain image segmentation. We generate feature sequences from brain images and feed them into a trained LSTM/BiLSTM model to obtain semantic labels. This method achieves promising segmentation results as well as robustness to noise.

Kai Xie, Ying Wen. LSTM-MA: A LSTM Method with Multi-modality and Adjacency Constraint for Brain Image Segmentation. (Submitted to ICIP 2019)

Matlab code for implemention of our method: LSTM-MA and BiLSTM-MA.

• BrainWeb: contains simulated MRI volumes for normal brain with three modalities: T1, T2 and PD.
• MRBrainS: contains T1, T1 inversion recovery and FLAIR sequences.

Illustration of our proposed segmentation pipeline. Given the input of multi-modality slices, two phases are followed to obtain the final segmentation result. First is the sequence construction phase, feature sequences are generated in two ways, namely pixel-wise constraint and superpixel-wise constraint. Second is the classification phase, feature sequences are fed into a LSTM or BiLSTM layer separately followed by fully connected and softmax layer.

Segmentation results for WM (white), GM (yellow), and CSF (red) in three orthogonal views on BrainWeb. We choose FCM, SVM, SegNet and BiLSTM-MA (ours) for comparison.

We test LSTM-MA and BiLSTM-MA with pixel-wise and superpixel-wise constraint on different levels of noise in comparison with other five segmentation methods, i.e., k-means, FCM, SVM, KNN, Decision Tree. BrainWeb provides image noise with Rayleigh noise. MRBrainS provides image noise with Gaussian noise.

For more information, please contact [email protected]