Unsupervised image registration based on Residual-connected DRMINE for diagnostic metastatic bone tumors

Metastatic tumors are frequently identified through follow-up surveillance using computed tomography (CT) scans.However, CT scans produce more than 100 images in an examination, which imposes a significant burden on radiologists and entails a potential risk of misdiagnosis.Temporal subtraction is utilized in Computer-Aided Diagnosis (CAD) and proves to be an effective technique in aiding image interpretation process for the radiologists.In this study, we focus on the preliminary stage of CAD development specialized in bone metastasis extraction, with a particular emphasis on rigid registration.We propose a novel rigid registration technique by augmenting DRMINE, which estimates mutual information using neural networks, with skip connections and normalization.From the three datasets, ten images were selected randomly from the cervical, thoracic, and lumbar regions.These images were then augmented with rotation as well as horizontal and vertical translations to create modified versions.The registration accuracy was assessed based on the Full Width at Half Maximum (FWHM) of the difference images.In the proposed method, FWHM values for the thoracic and lumbar regions of the spine exhibited a maximum reduction rate of 2.8% and a minimum reduction rate of 0.533%.However, the cervical spine region exhibited superior FWHM results with DRMINE compared to the proposed methodology.The proposed method was influenced by the capture area, but it indicated the potential to provide stable registration as the standard deviation decreased for all FWHM values.