Автоматична сегментація підшлункової залози з використанням ResNet-18 методу глибокого навчання

• Дата: 2022
• Автори: Kakarwal, Sangeeta, Paithane, Pradip
• Формат: Стаття
• Мова: English
• Опубліковано: The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" 2022
• Теми: глибоке навчання; коефіцієнт кубиків; повністю підключений шар (FCN); залишкова мережа (ResNet-18); група візуальної геометрії (VGG); Deep Learning; Dice Coefficient; Fully Connected Layer (FCN); Residual Network (ResNet-18); Visual Geometry Group (VGG)
• Онлайн доступ: http://journal.iasa.kpi.ua/article/view/265645
• Назва журналу: System research and information technologies

The accurate pancreas segmentation process is essential in the early detection of pancreatic cancer. The pancreas is situated in the abdominal cavity of the human body. The abdominal cavity contains the pancreas, liver, spleen, kidney, and adrenal glands. Sharp and smooth detection of the pancreas from this abdominal cavity is a challenging and tedious job in medical image investigation. Top-down approaches like Novel Modified K-means Fuzzy clustering algorithm (NMKFCM), Scale Invariant Feature Transform (SIFT), Kernel Density Estimator (KDE) algorithms were applied for pancreas segmentation in the early days. Recently, Bottom-up method has become popular for pancreas segmentation in medical image analysis and cancer diagnosis. LevelSet algorithm is used to detect the pancreas from the abdominal cavity. The deep learning, bottom-up approach performance is better than another. Deep Residual Network (ResNet-18) deep learning, bottom-up approach is used to detect accurate and sharp pancreas from CT scan medical images. 18 layers are used in the architecture of ResNet-18. The automatic pancreas and kidney segmentation is accurately extracted from CT scan images. The proposed method is applied to the medical CT scan images dataset of 82 patients. 699 images and 150 images with different angles are used for training and testing purposes, respectively. ResNet-18 attains a dice similarity index value up to 98.29±0.63, Jaccard Index value up to 96.63±01.25, Bfscore value up to 84.65±03.96. The validation accuracy of the proposed method is 97.01%, and the loss rate value achieves up to 0.0010. The class imbalance problem is solved by class weight and data augmentation.