Unsupervised learning-based deformable registration of temporal chest radiographs to detect interval change

Employing similarity to highlight differences: On the impact of anatomical assumptions in chest X-ray registration methods

“…The discriminator network penalizes, among standard terms, also the overlap of lung masks and additionally a cycle consistency loss is used. Small differences are also the subject of [11] and [14], which also start with affine and B-spline registration within segmented lung regions on a private data set, excluding cases with too large deformations. They penalize first and second derivatives of the dense displacement field during CNN training in order to obtain better local matching in an unsupervised setting.…”

“…At inference time, opposed to conventional methods, optimization is achieved in a single forward pass. Several authors have successfully investigated how to optimally warp the moving/source image to correspond with the fixed/target image based on surrogate measures for chest X-rays with the help of deep learning (see [1,11,29,30]). Although these methods are fast and often surpass classical methods in registration quality, they come with intrinsic challenges, concerning the technical implementations of producing smooth transformations and restricting or limiting folding.…”