Auto Segmentation Model for Head and Neck Cancer Patients Using U-Net with Convolutional Block Attention Modules
K Yamamoto1*, R Nakagami1, M Nakano2, H Ishiyama2, Y Tanaka3, T Hasegawa1, M Hashimoto1, (1) Kitasato University School of Allied Health Sciences, Sagamihara, Kanagawa, JP, (2) Kitasato University School of Medicine, Sagamihara, Kanagawa, JP, (3) Kitasato University Graduate School of Medical Sciences, Sagamihara, Kanagawa, JP
Presentations
PO-GePV-M-9 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: It is necessary to draw contours of numerous organs-at-risk (OARs) as well as treatment targets prior to radiotherapy planning of head and neck cancer patients. The aim of this study was to develop highly-precise segmentation model for OARs in head-and-neck region using Convolutional Neural Network with Attention Modules.
Methods: The developed network for head and neck organ segmentation was based on the structure of U-Net (ResNet-50 as backbone of encoder) and including Convolutional Block Attention Modules (CBAMs) in each skip connection and each layer in decoder part. CBAMs create attention maps for the input features in the channel and spatial directions, while maintaining the spatial information. The CT image sets of 139 head and neck cancer patients who were treated by intensity-modulated radiotherapy in Kitasato University Hospital were used in this study, as training (93), validation (24) and evaluation (22) data sets. The contours of organs were manually delineated by oncologists and used as ground truths of this study. The segmented organs are eyes, lens, optic nerves, chiasm, brainstem, parotid glands, submandibular glands, thyroid and spinal cord. Accuracy of the developed network was compared to the results of conventional U-Net using Dice indices.
Results: Overall average of Dice indices of the developed model was 1.1% improved compared to the conventional U-Net. Significant improvement was observed in lens, optic nerves and chiasm. Especially Dice indices of optic nerves were 0.7236 and 0.7644, by conventional U-Net and by the present model with CBAMs, respectively.
Conclusion: The developed system improved the accuracy of organ segmentation in head and neck region, especially in relatively small structures. These results indicated that the developed network with CBAMs was effective and robust for head and neck segmentation.
Keywords
Taxonomy
Contact Email