類神經網路應用於電阻抗斷層掃描

Abstract

電阻抗斷層掃描（EIT）是一個透過量測容器內的導體其電阻抗變化來成像的方法。電阻抗斷層掃描目前可以應用在工業、軍事、地質還有醫療等領域上。和其他醫療應用的技術比起來， 電阻抗斷層掃描是一個快速且便宜的方法，並且不會產生出可能會造成病人受病變的遊離輻射。然而，電阻抗斷層掃描有較低成像品質的缺點。隨著電阻抗斷層掃描的出現，很多重建影像的方法也被提出來。因為使用線性理論會使影像過於平滑，進一步的限制精準度。因此，出現了現代非線性理論應用在模擬上來產生出高品質的影像，但還是因為種種原因在醫療上未能產生令人滿意的結果。人工類神經網路（ANNs）被近年來視為有希望能夠使電阻抗斷層掃描成像更精確的非線性方法。這個方法不論在模擬或者是假體實驗上都獲得即佳的重建結果，但是要應用在醫療領域上人工類神經網路需要很完整的模型來做有效率的訓練。本論文呈現了用建立在人工類神經網路上的不同方法來提升電阻抗斷層掃描的成像品質並討論。在較小非均質物質的情況下，使用人工類神經網絡來估計非均質物質的位置可以顯著提升成像速度。因此，EIT的反向問題使用人工類神經網絡可以即時的解出來。而且，使用徑向基函數核（RBF）在沒有調整先前的轉換方程式任何參數下可以提供更高品質的影像。本論文介紹一個可以自動選取最佳的分散常數的方法來提升成像的品質。另外，一個建立在人工類神經網路理論上的新方法會被提出。這個方法的是結合了線性逆求解和非線性人工類神經網的優點並且消除了各自的缺點。這個後處理的方法有在二維影像以及三維影像側試過。使用真實的呼吸醫學影像可以證實這個方法同時有好的影像品質以及穩定性。

Electrical impedance tomography (EIT) is an imaging method based on the variations of electrical impedance within a volume conductor. EIT can be applied in various fields such as industry, military, geology and, last but not least, medical sector. Compared to other technologies actually used in biomedical applications, EIT offers the advantages of being a rapid and inexpensive method, and does not emit any ionizing radiations that are potentially harmful for the patient. However, the shortcomings of this technology lay in the low quality of the produced images. Since the idea of using electrical impedance as an imaging method came out, several methods have been proposed to do the image reconstruction. While linear methods usually generate large smoothness in the image, which limits the accuracy, modern nonlinear methods appear to give high resolution in simulation but still have troubles to give satisfactory solution in biomedical applications, for several reasons. Artificial Neural Networks (ANNs) are currently regarded as a promising method for accurate nonlinear EIT reconstruction. This method gives excellent result in simulations or phantom reconstructions, but their use in biomedical applications requires an important modelling effort in order to train them efficiently. In this dissertation, different methods to improve the quality of EIT images based on ANNs are presented and discussed. For the case of small inhomogeneities, the reconstruction speed can be significantly increased by estimating the location of the inhomogeneities using an ANN. Then, the EIT inverse problem can be solved in real time by using an ANN. In this case, the use of Radial Basis Function (RBF) is known to give high quality images, although previous work did not adjust the different parameters of the transfer functions. This dissertation introduces a way to automatically pick the best spread constant in the RBF ANN, which then enhances the quality of the reconstructions. Additionally, a novel method based on ANN is introduced. This method aims to combine the advantages of both linear inverse solvers and nonlinear ANNs, and to eliminate their respective drawbacks. This post-processing method is tested on both two-dimensional and three-dimensional imaging. The use of real biomedical breathing data confirms that this original method bring both quality and stability.