A novel deep learning architecture for multiple sclerosis Diagnosis : Gau-U-Net /
Roba Gamal Mohamed,
A novel deep learning architecture for multiple sclerosis Diagnosis : Gau-U-Net / بنية جديدة للتعلم العميق لتشخيص التصلب المتعدد : Gau-U-Net / by Roba Gamal Mohamed ; Under the Supervision of Prof. Hoda Baraka, Assoc. Prof. Mayada Mansour. - 96 pages : illustrations ; 30 cm. + CD.
Thesis (M.Sc.)-Cairo University, 2023.
Bibliography: pages 66-69.
Multiple sclerosis is an autoimmune disease that affects the brain and nervous system.
2.8 million people are estimated to live with MS worldwide (35.9 per 100,000
population). The pooled incidence rate across 75 reporting countries is 2.1 per 100,000
persons per year, and the mean age of diagnosis is 32 years. Lesions resulting from the
disease can be spotted in the patient’s MRI scans. In this paper, a novel Deep learning
architecture, GAU-U-net, is proposed. The model is inspired by the very famous U-Net
architecture used for semantic segmentation and is frequently employed in the
segmentation of medical images. The proposed model consists of a
3D U-Net after adding a new attention technique inspired by the Global Attention
Upsample (GAU) unit. After using GAU-unet architecture, the Dice coefficient increased
from 64% to 72% compared to 3D-Unet.Also, compared with the Unet attention network,
the dice coefficient increased from 69% to around 72%, with a considerable decline in
the number of model parameters in favor of our architecture, which uses 28M parameters
compared to Unet-attention, which employs 100 M parameters ﯾﺘﺄﺛﺮ اﻟﺪﻣﺎغ واﻟﺤﺒﻞ اﻟﺸﻮﻛﻲ ﺑﻤﺮض اﻟﺘﺼﻠﺐ اﻟﻤﺘﻌﺪد، وھﻮ ﻣﻦ اﺿﻄﺮاﺑﺎت اﻟﻤﻨﺎﻋﺔ اﻟﺬاﺗﯿﺔ. وﻋﺎدة ﻣﺎ ﯾﺘﻢ ﺗﺸﺨﯿﺼﮫ ﻣﻦ ﺧﻼل اﻟﺘﺼﻮﯾﺮ ﺑﺎﻟﺮﻧﯿﻦ اﻟﻤﻐﻨﺎطﯿﺴﻲ.ﯾﺘﻄﻮر اﻟﻤﺮض إﻟﻰ آﻓﺎت ﺑﯿﻀﺎء ﯾﻤﻜﻦ أن ﺗﻜﻮن واﺿﺤﺔ
ﻓﻲ اﻟﺘﺼﻮﯾﺮ ﺑﺎﻟﺮﻧﯿﻦ اﻟﻤﻐﻨﺎطﯿﺴﻲ وﯾﻤﻜﻦ ﺗﻘﺴﯿﻢ اﻵﻓﺎت ﯾﺪوﯾًﺎ ﺑﻮاﺳﻄﺔ اﻟﺨﺒﺮاء، وھﺬا ﯾﺴﺘﻐﺮق وﻗﺘًﺎ طﻮﯾﻼً
ﻟﻠﻐﺎﯾﺔ.إن ﺗﻄﺒﯿﻖ ﺗﻘﻨﯿﺎت اﻟﺮؤﯾﺔ اﻟﺤﺎﺳﻮﺑﯿﺔ ﻟﺘﺠﺰﺋﺔ اﻵﻓﺎت ﯾﻤﻜﻦ أن ﯾﺠﻌﻞ اﻟﻌﻤﻠﯿﺔ أﺳﮭﻞ وأﺳﺮع، ﻓﻲ ھﺬه
ﻣﻘﺒﻮل ﻣﻦ
ﻣﻊ ﻋﺪد
أﻋﻠﻰ
دﻗﺔ
ﻟﺘﺤﻘﯿﻖ
اﻻﻧﺘﺒﺎه
آﻟﯿﺔ
ﺗﺴﺘﻐﻞ
ﺟﺪﯾﺪة
دﻻﻟﯿﺔ
ﺗﺠﺰﺋﺔ
ﺑﻨﯿﺔ
ھﻨﺎك
اﻷطﺮوﺣﺔ
اﻟﻤﺘﻐﯿﺮات. ﯾﺘﻢ اﺳﺘﺨﺪام ﻧﻔﺲ اﻟﺒﻨﯿﺔ اﻟﺠﺪﯾﺪة ﻋﻠﻰ ﺑﯿﺎﻧﺎت ﺛﻨﺎﺋﯿﺔ اﻷﺑﻌﺎد ﺑﻤﺴﺎﻋﺪة اﻟﺘﻌﻠﻢ اﻟﻤﺘﻨﻘﻞ وھﻮ ﻣﺎ ﯾﻮﺿﺢ
اﻟﻤﺠﺎﻻت اﻟﺘﻲ ﯾﺼﻌﺐ اﻟﺤﺼﻮل ﻋﻠﻰ اﻟﺒﯿﺎﻧﺎت ﻓﯿﮭﺎ ﻛﻤﺎ ھﻮ اﻟﺤﺎل ﻓﻲ
اﻟﻤﺘﻨﻘﻞ ﻓﻲ
اﻟﺘﻌﻠﻢ
اﺳﺘﺨﺪام
ﺗﺄﺛﯿﺮ
اﻟﻤﺠﺎل اﻟﻄﺒﻲ .ارﺗﻔﻊ ﻣﻌﺎﻣﻞ Dice ﻣﻦ %64 إﻟﻰ %72 ﺑﺎﻟﻤﻘﺎرﻧﺔ إﻟﻰ 3D-Unet ﺑﻌﺪ اﺳﺘﺨﺪام اﻟﺒﻨﯿﺔ اﻟﻤﻘﺘﺮﺣﺔ ﻓﻲ اﻟﺒﺤﺚ .ﺑﺎﻹﺿﺎﻓﺔ إﻟﻰ ذﻟﻚ، ﻋﻨﺪ اﻟﻤﻘﺎرﻧﺔ ﺑﺎﻟﻨﺴﺒﺔ إﻟﻰ ﺷﺒﻜﺔ Unet-attention، زاد ﻣﻌﺎﻣﻞ Dice ﻟﺼﺎﻟﺢ اﻟﺒﻨﯿﺔ اﻟﻤﻘﺘﺮﺣﺔ ﻓﻲ اﻟﺒﺤﺚ ﻣﻦ 69 إﻟﻰ %72 ﺗﻘﺮﯾﺒًﺎ، ﻣﻊ اﻧﺨﻔﺎض ﻛﺒﯿﺮ ﻓﻲ ﻋﺪد
اﻟﻤﺘﻐﯿﺮات ﻣﻘﺎرﻧﺔ ﺑﺎﻟﻤﺘﻐﯿﺮات ﻓﻲ ﺑﻨﯿﺔ Unet-attention اﻟﺒﺎﻟﻐﺔ 100 ﻣﻠﯿﻮن.
Text in English and abstract in Arabic & English.
Computer Engineering
Multiple sclerosis lesion segmentation Deep learning Computer aided diagnosis Unet Semantic segmentation Attention Global attention Upsample
621.39
A novel deep learning architecture for multiple sclerosis Diagnosis : Gau-U-Net / بنية جديدة للتعلم العميق لتشخيص التصلب المتعدد : Gau-U-Net / by Roba Gamal Mohamed ; Under the Supervision of Prof. Hoda Baraka, Assoc. Prof. Mayada Mansour. - 96 pages : illustrations ; 30 cm. + CD.
Thesis (M.Sc.)-Cairo University, 2023.
Bibliography: pages 66-69.
Multiple sclerosis is an autoimmune disease that affects the brain and nervous system.
2.8 million people are estimated to live with MS worldwide (35.9 per 100,000
population). The pooled incidence rate across 75 reporting countries is 2.1 per 100,000
persons per year, and the mean age of diagnosis is 32 years. Lesions resulting from the
disease can be spotted in the patient’s MRI scans. In this paper, a novel Deep learning
architecture, GAU-U-net, is proposed. The model is inspired by the very famous U-Net
architecture used for semantic segmentation and is frequently employed in the
segmentation of medical images. The proposed model consists of a
3D U-Net after adding a new attention technique inspired by the Global Attention
Upsample (GAU) unit. After using GAU-unet architecture, the Dice coefficient increased
from 64% to 72% compared to 3D-Unet.Also, compared with the Unet attention network,
the dice coefficient increased from 69% to around 72%, with a considerable decline in
the number of model parameters in favor of our architecture, which uses 28M parameters
compared to Unet-attention, which employs 100 M parameters ﯾﺘﺄﺛﺮ اﻟﺪﻣﺎغ واﻟﺤﺒﻞ اﻟﺸﻮﻛﻲ ﺑﻤﺮض اﻟﺘﺼﻠﺐ اﻟﻤﺘﻌﺪد، وھﻮ ﻣﻦ اﺿﻄﺮاﺑﺎت اﻟﻤﻨﺎﻋﺔ اﻟﺬاﺗﯿﺔ. وﻋﺎدة ﻣﺎ ﯾﺘﻢ ﺗﺸﺨﯿﺼﮫ ﻣﻦ ﺧﻼل اﻟﺘﺼﻮﯾﺮ ﺑﺎﻟﺮﻧﯿﻦ اﻟﻤﻐﻨﺎطﯿﺴﻲ.ﯾﺘﻄﻮر اﻟﻤﺮض إﻟﻰ آﻓﺎت ﺑﯿﻀﺎء ﯾﻤﻜﻦ أن ﺗﻜﻮن واﺿﺤﺔ
ﻓﻲ اﻟﺘﺼﻮﯾﺮ ﺑﺎﻟﺮﻧﯿﻦ اﻟﻤﻐﻨﺎطﯿﺴﻲ وﯾﻤﻜﻦ ﺗﻘﺴﯿﻢ اﻵﻓﺎت ﯾﺪوﯾًﺎ ﺑﻮاﺳﻄﺔ اﻟﺨﺒﺮاء، وھﺬا ﯾﺴﺘﻐﺮق وﻗﺘًﺎ طﻮﯾﻼً
ﻟﻠﻐﺎﯾﺔ.إن ﺗﻄﺒﯿﻖ ﺗﻘﻨﯿﺎت اﻟﺮؤﯾﺔ اﻟﺤﺎﺳﻮﺑﯿﺔ ﻟﺘﺠﺰﺋﺔ اﻵﻓﺎت ﯾﻤﻜﻦ أن ﯾﺠﻌﻞ اﻟﻌﻤﻠﯿﺔ أﺳﮭﻞ وأﺳﺮع، ﻓﻲ ھﺬه
ﻣﻘﺒﻮل ﻣﻦ
ﻣﻊ ﻋﺪد
أﻋﻠﻰ
دﻗﺔ
ﻟﺘﺤﻘﯿﻖ
اﻻﻧﺘﺒﺎه
آﻟﯿﺔ
ﺗﺴﺘﻐﻞ
ﺟﺪﯾﺪة
دﻻﻟﯿﺔ
ﺗﺠﺰﺋﺔ
ﺑﻨﯿﺔ
ھﻨﺎك
اﻷطﺮوﺣﺔ
اﻟﻤﺘﻐﯿﺮات. ﯾﺘﻢ اﺳﺘﺨﺪام ﻧﻔﺲ اﻟﺒﻨﯿﺔ اﻟﺠﺪﯾﺪة ﻋﻠﻰ ﺑﯿﺎﻧﺎت ﺛﻨﺎﺋﯿﺔ اﻷﺑﻌﺎد ﺑﻤﺴﺎﻋﺪة اﻟﺘﻌﻠﻢ اﻟﻤﺘﻨﻘﻞ وھﻮ ﻣﺎ ﯾﻮﺿﺢ
اﻟﻤﺠﺎﻻت اﻟﺘﻲ ﯾﺼﻌﺐ اﻟﺤﺼﻮل ﻋﻠﻰ اﻟﺒﯿﺎﻧﺎت ﻓﯿﮭﺎ ﻛﻤﺎ ھﻮ اﻟﺤﺎل ﻓﻲ
اﻟﻤﺘﻨﻘﻞ ﻓﻲ
اﻟﺘﻌﻠﻢ
اﺳﺘﺨﺪام
ﺗﺄﺛﯿﺮ
اﻟﻤﺠﺎل اﻟﻄﺒﻲ .ارﺗﻔﻊ ﻣﻌﺎﻣﻞ Dice ﻣﻦ %64 إﻟﻰ %72 ﺑﺎﻟﻤﻘﺎرﻧﺔ إﻟﻰ 3D-Unet ﺑﻌﺪ اﺳﺘﺨﺪام اﻟﺒﻨﯿﺔ اﻟﻤﻘﺘﺮﺣﺔ ﻓﻲ اﻟﺒﺤﺚ .ﺑﺎﻹﺿﺎﻓﺔ إﻟﻰ ذﻟﻚ، ﻋﻨﺪ اﻟﻤﻘﺎرﻧﺔ ﺑﺎﻟﻨﺴﺒﺔ إﻟﻰ ﺷﺒﻜﺔ Unet-attention، زاد ﻣﻌﺎﻣﻞ Dice ﻟﺼﺎﻟﺢ اﻟﺒﻨﯿﺔ اﻟﻤﻘﺘﺮﺣﺔ ﻓﻲ اﻟﺒﺤﺚ ﻣﻦ 69 إﻟﻰ %72 ﺗﻘﺮﯾﺒًﺎ، ﻣﻊ اﻧﺨﻔﺎض ﻛﺒﯿﺮ ﻓﻲ ﻋﺪد
اﻟﻤﺘﻐﯿﺮات ﻣﻘﺎرﻧﺔ ﺑﺎﻟﻤﺘﻐﯿﺮات ﻓﻲ ﺑﻨﯿﺔ Unet-attention اﻟﺒﺎﻟﻐﺔ 100 ﻣﻠﯿﻮن.
Text in English and abstract in Arabic & English.
Computer Engineering
Multiple sclerosis lesion segmentation Deep learning Computer aided diagnosis Unet Semantic segmentation Attention Global attention Upsample
621.39