以聯邦學習提升少量資料之腦腫瘤切割模型訓練成效

Abstract

近年來隨著深度神經網路的蓬勃發展，也越來越多使用於醫學影像的相關研究，其中腦腫瘤分割模型的訓練正是其一。不過，很多醫療院所礙於資料集的數量小，沒有辦法達到最佳化的模型訓練，容易有Overfitting等等的問題，然而資料集共享又會衍生出許多像是病人隱私、合法性及資料擁有權的問題。在這樣的前提下聯邦學習的訓練模式便是極佳的選擇。 聯邦學習是一種能夠同時保障資料安全性，並達到分散式學習的方法。在聯邦學習的學習過程中，模型會由伺服器端送至客戶端進行訓練，每訓練一個節點，客戶端便會回傳模型參數回伺服器端，讓伺服器可以進行模型調整，再送回客戶端持續進行訓練。在訓練過程中同時達到資料不共享，但資料量加成提供模型學習的優勢，便是聯邦學習崛起的起因。 而我們該如何運用聯邦學習的模式來幫助醫院之間，在資料不共享的前提下幫助腦瘤分割模型的訓練發展，使資料量不足的醫療院所也可以訓練出良好的腦瘤分割模型就成了很重要的任務。 本篇論文使用了Clara SDK 4.0版本作為開發的環境，也運用了MONAI來作為深度學習應用於醫學影像訓練的框架。在針對兩間醫院的數據集訓練過後，發現進行聯邦學習之後可以大幅降低資料集小的醫院訓練結果的擬合過度，也提升了訓練的結果的Dice係數。

More and more research about medical images recently arose, with deep neuron networks thriving. Research about brain tumor segmentation was one of them. But many medical institutions can’t reach the best model training due to the lack of data. It’s easy to get problems with overfitting, etc. However, the sharing of datasets is a significant challenge due to patients' privacy, legal issues, and also data-ownership issues. To face such issues, Federated Learning is one of the best remedies. Federated learning is a way of training that can ensure data security and reach distributed learning spontaneously. The model will be transferred from the server-side to the client-side during federated learning. Each time a node is trained, the client will send the model parameters back to the server. Then the server will adjust the model and send it back to the client for continuous training. In the training process, data are not shared at the same time. But the increased data volume provides the advantage of model learning, which is why federated learning is rising. It is an essential task for us to see how we can use federated learning to improve the brain tumor segmentation model development without sharing data in medical institutions that lack data so that those institutions can also train a robust brain tumor segmentation model. In the thesis, Clara SDK 4.0 is used for the development environment. Besides, MONAI (Medical Open Network for AI) is used as a deep learning frame in medical image training. After training on the datasets of two hospitals, it was found that federated learning can significantly reduce the overfitting of the training results of hospitals with small datasets. In addition, it can also improve the dice score of the training results.