قطعه بندي تومور ضايعه مغزي با استفاده از شبكه عصبي كانولوشن

Brain Tumor Segmentation Using the Convolutional Neural Network

در اﯾﻦ ﻣﻘﺎﻟﻪ روش ﺟﺪﯾﺪي ﻣﺒﺘﻨﯽ ﺑﺮ ﺷﺒﮑﻪ ﻫﺎي ﻋﺼﺒﯽ ﮐﺎﻧﻮﻟﻮﺷﻦ ﺑﺮاي ﻗﻄﻌﻪ ﺑﻨﺪي ﺗﻮﻣﻮرﻫﺎي ﻣﻐﺰي در ﺗﺼﺎوﯾﺮ ام آر آي اراﺋﻪ ﻣﯽ ﺷﻮد ﮐﻪ ﮐﺎر را ﺑﺎ ﻣﺮﺣﻠﻪ ﭘﯿﺶ ﭘﺮدازش ﺷﺎﻣﻞ اﺻﻼح ﻣﯿﺪان ﺑﺎﯾﺎس، ﺷﺪت و ﻋﺎدي ﺳﺎزي ﻗﻄﻌﺎت ﺷﺮوع ﻣﯽ ﺷﻮد ﭘﺲ از آن ﻃﯽ روﻧﺪ آﻣﻮزش، ﺗﻌﺪاد ﻗﻄﻌﺎت آﻣﻮزﺷﯽ ﺑﻪ ﻃﻮر ﻣﺼﻨﻮﻋﯽ ﺑﻪ ﮐﻤﮏ ﭼﺮﺧﺶ ﻗﻄﻌﺎت آﻣﻮزﺷﯽ و ﻧﯿﺰ اﺳﺘﻔﺎده از ﻧﻤﻮﻧﻪ ﻫﺎي HGG ﺑﺮاي ﻗﻄﻌﻪ ﺑﻨﺪي ﺗﻌﺪاد دﺳﺘﻪ ﻫﺎي ﮐﻤﯿﺎب LGG اﻓﺰاﯾﺶ ﻣﯽ ﯾﺎﺑﺪ. ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﮐﺎﻧﻮﻟﻮﺷﻦ روي ﻻﯾﻪ ﻫﺎي ﮐﺎﻧﻮﻟﻮﺷﻦ ﺑﺎ اﺑﻌﺎد ﻫﺴﺘﻪ ﮐﻮﭼﮏ 3 × 3 ﺳﺎﺧﺘﻪ ﻣﯽ ﺷﻮد ﺗﺎ اﻣﮑﺎن اﻧﺠﺎم ﻣﻌﻤﺎري ﻫﺎي ﻋﻤﯿﻖ ﺗﺮ را ﻓﺮاﻫﻢ ﻧﻤﺎﯾﺪ. در ﭘﺎﯾﺎن ﻧﺸﺎن داده ﻣﯽ ﺷﻮد ﮐﻪ ﺗﺎﺑﻊ ﻓﻌﺎل ﺳﺎزي LReLU ﻧﺴﺒﺖ ﺑﻪ ﺗﺎﺑﻊ ReLU ﻧﻘﺶ ﻣﻬﻢ ﺗﺮي در آﻣﻮزش ﮐﺎرآﻣﺪ ﺷﺒﮑﻪ ﻫﺎي ﻋﺼﺒﯽ ﮐﺎﻧﻮﻟﻮﺷﻦ دارد.

This paper presents a new method based on convolutional neural networks for brain tumor segmentation in MRI images, which begins with the preprocessing stage, including the correction of the bias field, the intensity and the normalization of the parts. Then, during the training process, the numbers of training parts are increased artificially with the help of rotation of training parts and the use of HGG samples for segmentation of the small number of LGGs. The convolutional neural network is constructed on convolutional layers with the small core dimensions of 3 × 3 to allow for deeper architectures. Finally, it is shown that the activation function of LReLU plays a more important role in the efficient training of convolutional neural networks than the ReLU function.