基於非監督式學習之無交疊B-Spline醫學影像配準

Abstract

非剛體影像配準用來建立一組移動（moving）及模板（fixed）影像間， 像素對像素的對應關係，來對齊不同影向間的資訊，目前在醫學影像分析 上已經有廣泛的應用。傳統醫學影像配準技術，將這個目標視作最佳化問 題，透過重複迭代逐步求解，然而這個過程十分耗費時間與運算資源，也限 制了這項技術在實務上許多的應用。近期提出的方法，使用深度網路學習， 預先訓練一組用於影像配準的參數，直接預測將影像組配準所需的形變場 （deformation field)，大大縮短的計算所需的時間。然而，直接預測每個像 素的位移向量，可能會產生不可能在生物體中所產生的變形。因此這篇研 究，實現了類神經卷積網路的 B-spline 變換，來確保產生的變換場是平滑且 連續的，另外也提出 folding-free loss 來解決結果中折疊的情況發生。另外 透過多尺度學習，讓網路同時考慮不同尺度下的特徵，可以讓我們的方法在 ACDC 心臟 MRI 影像的配準準確度上，達到當前最優的結果。

Deformable image registration, which establishes a non-linear correspondence between a pair of images, is widly use a fundamental step in many medical image analysis procedures. Conventional registration methods, which iteratively solve an optimization problem, can be very slow and hinder the practical applications. To this end, formulating it as a learning based problem using Convolutional Neural Networks (CNNs) can largely reduce the registration time [3]. Most of existing learning-based methods directly predict a dense deformation field from an input pair of fixed and moving images. However, the resulting transformation could be physically implausible, and selffolding cannot be avoided. In this work, we model the deformation with Bspline transformation which intrinsically produces smooth deformation, and the proposed local invertibility constraint largely alleviates the self-folding issue in the resulting deformation field. We also present that our multi-scale learning framework can further improve the registration accuracy. The proposed approach is trained in an unsupervised fashion, and no ground-truth registration fields or landmark annotations are needed. Experimental results demonstrate that our registration method outperforms current state-of-the-art algorithms using public available ACDC cardiac cine MRI dataset [25].