Liam Chalcroft

LKA: Large-kernel Attention for Efficient and Robust Brain Lesion Segmentation

L. Chalcroft, R.L. Pereira, M. Brudfors, A.S. Kayser, M. D'Esposito, C.J. Price, J. Ashburner
Medical Imaging meets NeurIPS 2023 Workshop

Vision transformers are effective deep learning models for vision tasks, including medical image segmentation. However, they lack efficiency and translational invariance, unlike convolutional neural networks (CNNs). To model long-range interactions in 3D brain lesion segmentation, we propose an all-convolutional transformer block variant of the U-Net architecture. We demonstrate that our model provides the greatest compromise in three factors: performance competitive with the state-of-the-art; parameter efficiency of a CNN; and the favourable inductive biases of a transformer.

Block diagrams of the LKA architecture. Three sub-blocks are shown: LKA (depthwise, dilated depthwise and pointwise convolutions combined multiplicatively), attention (pointwise convolution, GELU, LKA, pointwise convolution with a residual connection), and a convolutional feed-forward block. Below, these compose into the full LKA block: overlapping patch embedding, then N repeats of batch norm plus attention and batch norm plus feed-forward, each with residual connections, ending in layer norm.
Cite as

Chalcroft, L.F., et al. (2023). LKA: Large-kernel Attention for Efficient and Robust Brain Lesion Segmentation. In Medical Imaging meets NeurIPS 2023 Workshop.