使用具有邊界損失的從粗到細神經結構搜索進行三維腦轉移瘤分割

Abstract

從圖像註釋標籤對於醫生在醫學圖像問題的工作流程中是很費時的。因此，我們的目標是建立一個自動分割流程，以幫助醫生進行註釋。整體工作流程分為預處理，模型訓練，後處理和驗證。對於模型的結構，我們使用由NVIDIA使用神經體系結構搜索（NAS）構建的粗到細神經結構搜索（C2FNAS）模型。在本文中，我們將焦點骰子損失與邊界損失相結合並比較性能，然後通過實驗觀察到，與僅使用焦點骰子損失的情況相比，邊界損失使訓練收斂更快，並且獲得了更好的性能。最後，我們從臺灣大學附屬醫院（NTUH）的數據集中獲得了最好的平均骰子得分為0.81。

Annotating the label from the images is time-consuming for doctors in the workflow of the medical image problem. Thus, we aim to build an automatic segmentation workflow to help doctor for annotation. The workflow is divided into pre-processing, model training, post-processing, and validation.For the architecture of model, we use the Coarse-to-Fine Neural ArchitectureSearch (C2FNAS) model builded by NVIDIA using the Neural ArchitectureSearch (NAS). In this thesis, we combine focal dice loss with boundary loss to compare performance, and then we observe by experiments that boundary loss make the convergence of training faster and get the better performance than the condition which only uses focal dice loss. Finally, we get the mean dice score 0.81 in our dataset which is from the National Taiwan UniversityHospital (NTUH).