Caln: Channel Attention LSTM Network for Pituitary Segmentation in Dynamic Contrast-Enhanced MRI

Automatic pituitary segmentation in MR images is crucial for pituitary-related diseases diagnosis, yet it is challenging due to limited spatial and temporal resolutions of the image and the small size of the gland. In this study, we introduce the Channel Attention Long Short-Term Memory Network (CALN), integrating Long Short-Term Memory (LSTM) units into the bottleneck of a symmetrical encoder-decoder structure. We employ channel-wise attention to fuse features extracted from different time phases based on their semantic importance. Notably, CALN achieves remarkable performance with only two downsampling layers. To evaluate the effectiveness of CALN, we conducted extensive experiments and compared it with state-of-the-art methods for pituitary segmentation. Our results show that CALN outperforms UNETR by 3.2% and SwinUNETR by 1.2% in terms of the dice score. The findings confirm the improved accuracy and computational efficiency of our method for pituitary segmentation.