Multiresolution Encoder-Decoder Convolutional Neural Network for Magnetic Resonance Image Segmentation

In image segmentation by deep learning, encoder-decoder Convolutional Neural Network (CNN) architectures are fundamental for creating and learning representations. However, with many filters in these models, one often loses the tractability of these models and interprets them as black boxes. In this paper, we describe the design of a multiresolution encoder-decoder CNN having a distinct low-frequency processing encoder and mixed-frequency processing decoder. The encoder of the proposed architecture first creates a localized spatial-frequency representation of the feature space using a framelet system and then learns representations using trainable convolution layers while converging to segment structures. The decoder takes in the encoder output, including distinct low and high-frequency skip connections and reconstructs the Region of Interest (RoI) with accurate boundaries. Spatial attention layers for low frequencies and a wavelet-based self-attention for high frequencies in the decoder successively aid the reconstruction of the desired segment. The described network architecture is overall lossless by design, despite information loss in wavelet pooling layers because of the initial subband decomposition and separate low and high-frequency paths. The model is tested separately using T1-weighted brain Magnetic Resonance (MR) images with two classes and accurately segments large and small RoIs. The trained model also successfully segments MR volumes acquired by our group.