Open AccessSynthesising KV‐DRRs from MV‐DRs with fractal hourglass convolutional network
Author(s) -
Liu Cong,
Huang Miao,
Ma Longhua,
Lu Zheming
Publication year - 2018
Publication title -
electronics letters
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.375
H-Index - 146
eISSN - 1350-911X
pISSN - 0013-5194
DOI - 10.1049/el.2017.4572
Subject(s) - hourglass , computer science , artificial intelligence , convolutional neural network , computer vision , fractal , graphics , flops , key (lock) , computer graphics (images) , mathematics , physics , mathematical analysis , computer security , astronomy , parallel computing
During a radiation treatment, the images of kilovoltage digital reconstructed radiograph (KV‐DRR) and megavoltage digital radiograph (MV‐DR) are registered to guide the therapy. Such registration is difficult since the images belong to different modalities. To reduce the difficulty, a fractal convolutional network is developed to map MV‐DR images into the modality of KV‐DRRs. The key idea is to split a hourglass‐shape network into multiple similar networks at reduced scale, yielding a fractal topology that is self‐similar at multiple scales. This division allows to predict images of unprecedented high resolution at low graphics processing unit memory usage. Experiments demonstrate that perceptual plausible and numerical accurate results are achieved out‐competing recent alternative architectures.