Abstract— Accurately reconstructing 3-D hand shapes ofpatients is important for immobilization device customization,artificial limb generation, and hand disease diagnosis. Traditional3-D hand scanning requires multiple scans taken around thehand with a 3-D scanning device. These methods require thepatients to keep an open-palm posture during scanning, whichis painful or even impossible for patients with impaired handfunctions. Once multi-view partial point clouds are collected,expensive post-processing is necessary to generate a high-fidelityhand shape. To address these limitations, we propose a noveldeep-learning method dubbed PatientHandNet to reconstructhigh-fidelity hand shapes in a canonical open-palm pose frommultiple-depth images acquired with a single-depth camera. Thehand poses in the depth images may vary, hand movementsare allowed, facilitating the 3-D scanning process in particularfor patients with difficult conditions. The proposed methodhas strong operability since it is insensitive to the input pose,allowing for pose variations in the input depth images. We alsoproposed two novel datasets: a large-scale synthetic dataset totrain our model and a real-world dataset with ground-truthhand biometrics extracted by an experienced anthropometrist.Extensive experimental results on the unseen synthetic dataand real-world data demonstrate that the proposed methodprovides robust and easy-to-use hand shape reconstruction andoutperforms the state-of-the-art methods in biometric accuracyterms.