Slimmable transformer with hybrid axial-attention for medical image segmentation

The transformer architecture has achieved remarkable success in medical image analysis owing to its powerful capability for capturing long-range dependencies. However, due to the lack of intrinsic inductive bias in modeling visual structural information, the transformer generally requires a large-scale pre-training schedule, limiting the clinical applications over expensive small-scale medical data. To this end, we propose a slimmable transformer to explore intrinsic inductive bias via position information for medical image segmentation. Specifically, we empirically investigate how different position encoding strategies affect the prediction quality of the region of interest (ROI) and observe that ROIs are sensitive to different position encoding strategies. Motivated by this, we present a novel Hybrid Axial-Attention (HAA) that can be equipped with pixel-level spatial structure and relative position information as inductive bias. Moreover, we introduce a gating mechanism to achieve efficient feature selection and further improve the representation quality over small-scale datasets. Experiments on LGG and COVID-19 datasets prove the superiority of our method over the baseline and previous works. Internal workflow visualization with interpretability is conducted to validate our success better; the proposed slimmable transformer has the potential to be further developed into a visual software tool for improving computer-aided lesion diagnosis and treatment planning.

Magnetic resonance imaging of mechanical deformations

A method for magnetic resonance imaging of mechanical deformations is presented. The method utilizes an MRI compatible device for inducing elastic deformations of a sample and a modified spin-echo imaging sequence with two position-encoding gradients added to the sequence symmetrically to the RF refocusing pulse. At the end of the first position-encoding gradient pulse, a sample deformation was induced by the deformational device, which applied a force to a plastic rod embedded in a gelatin cylindrical sample. The sample had to withstand repeated elastic deformations. Sample displacements up to 400 μm were encoded in the image signal phase by the use of position-encoding gradients. Images of different displacement components were acquired first by the use of position-encoding gradients in different directions and then processed by the 2D phase unwrap algorithm. Finally, images of normal and shear strain distribution were calculated from the displacement images. The obtained displacement and strain images enabled clear visualization of deformations and their extent in the sample with the displacement detection threshold in the range 0.3-0.6 μm, depending on the image echo time. The results of displacements were verified also by a DANTE tagging method and by an optical method. The presented method enables studying of various types of deformations in different soft materials as well as dynamic response of deformations to different stress functions (static, oscillatory, pulsed…).