We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

eess.IV

Change to browse by:

References & Citations

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Electrical Engineering and Systems Science > Image and Video Processing

Title: Multi-structure segmentation for renal cancer treatment with modified nn-UNet

Authors: Zhenyu Bu
Abstract: Renal cancer is one of the most prevalent cancers worldwide. Clinical signs of kidney cancer include hematuria and low back discomfort, which are quite distressing to the patient. Due to the rapid growth of artificial intelligence and deep learning, medical image segmentation has evolved dramatically over the past few years. In this paper, we propose modified nn-UNet for kidney multi-structure segmentation. Our solution is founded on the thriving nn-UNet architecture using 3D full resolution U-net. Firstly, various hyperparameters are modified for this particular task. Then, by doubling the number of filters in 3D full resolution nnUNet architecture to achieve a larger network, we may capture a greater receptive field. Finally, we include an axial attention mechanism in the decoder, which can obtain global information during the decoding stage to prevent the loss of local knowledge. Our modified nn-UNet achieves state-of-the-art performance on the KiPA2022 dataset when compared to conventional approaches such as 3D U-Net, MNet, etc.
Comments: KiPA2022 Challenge report
Subjects: Image and Video Processing (eess.IV); Computer Vision and Pattern Recognition (cs.CV)
Cite as: arXiv:2208.05241 [eess.IV]
  (or arXiv:2208.05241v1 [eess.IV] for this version)

Submission history

From: Zhenyu Bu [view email]
[v1] Wed, 10 Aug 2022 09:49:19 GMT (1241kb,D)

Link back to: arXiv, form interface, contact.