基於深度學習之染色神經影像表皮層組織範圍分割

Abstract

神經系統是人類體內的重要系統，負責傳遞訊息、調節身體功能與維持內部平衡等重要生命運作功能。該系統可以分為兩個部分，分別為由脊椎神經與腦神 經構成之中樞神經系統，以及中樞神經系統以外的神經構成之周邊神經系統。其中周邊神經系統之病變通常會藉由染色神經影像進行診斷，唯診斷時使用人工分 割表皮層組織範圍需要耗費大量人工成本。本研究提出一種基於深度學習自動分割神經影像中表皮層組織範圍之模型，其由不同深度的U型網路以及多尺度模塊所構成，並且提出相應之前處理演算法，輔助本研究提出模型之驗證，以有效進行訓練與測試。本研究所提出之模型架構經過訓練後的預測表皮層組織範圍成效，可以在測試資料集展現優異之分割結果，並在Dice指標顯示具有92.80%之分割準確度。

The nervous system is an important system for the living body, responsible for transmitting information, regulating body functions, and maintaining internal balance and other important life functions. The system can be divided into two parts: the central nervous system composed of the spinal nerves and cranial nerves and the peripheral nervous system (PNS) composed of nerves other than the central nervous system. Among them, the diseases of the PNS are usually diagnosed by stained nerve images. But the segmentation of the epidermis area in the images requires a lot of labor costs. This thesis proposes a model based on deep learning to automatically segment the epidermis area in stained nerve images. The model is composed of U-shaped networks of different depths and multi-scale modules. We also propose pre-processing algorithms to facilitate the training and testing of the proposed model. The model architecture proposed in this study to predict the epidermis area shows excellent segmentation results with the Dice index indicating segmentation accuracy of 92.80%.