Small Lesion Segmentation in Brain MRIs with Subpixel Embedding

HUT: Hybrid UNet transformer for brain lesion and tumour segmentation

“…As shown in Table 6 , our HUT method improves the mean Dice score performance over the state-of-the-art SPiN [5] architecture by 4.84%. It also improves the mean 95th percentile Hausdorff Distance score (HD95) by 41%, as shown in the same table.…”

“…We cropped each 3D scan to a resolution of to focus on relevant portions of the image and reduce the GPU memory needed. To compare with the results in [5] , we used the same random data split of the ATLAS dataset the authors had evaluated, with 212 train and 27 test subjects. The ischemic stroke dataset contains very small lesions, which can make segmentation tasks difficult.…”

“…Wong et al [5] qualified anomalies less than 100 pixels as small lesions. Similarly, we also evaluated the performance of segmenting small lesions.…”

“…It also performs significantly better than Unetr on the ATLASR12 dataset, with a 16.0% improvement in the dice score and a 55.2% improvement in HD95. All other methods described in [5] performed worse than HUT and are shown in Table 6 . The same results for the task of small lesion segmentation are compared in Table 7 , where HUT outperforms all other methods by a significant margin.…”

“…As for the task of lesion segmentation, [5] demonstrated an approach to achieve state-of-the-art segmentation for the ATLAS R1.2 dataset. Their system uses two networks to make predictions.…”