2026-10-13T21:29:28Z https://www.scidb.cn/oai

10.57760/sciencedb.012dj 2026-09-23T19:48:51Z

2026-09-23 HAO-Former: A General Multi-Joint Musculoskeletal MRI Intelligent Denoising and Microstructure-Preserving Enhancement Platform (v1.0) doi:10.57760/sciencedb.012dj en Musculoskeletal magnetic resonance imaging (MSK-MRI) often suffers from pronounced stochastic thermal noise during rapid routine scans, while conventional deep-learning denoising networks frequently introduce over-smoothing ("wax-like" artifacts), eroding crucial microstructures such as trabecular bone networks, articular cartilage interfaces, and tendon fascicles.HAO-Former is a specialized, open-source medical software platform designed for intelligent denoising and diagnostic image enhancement across multi-joint MSK-MRI examinations. Powered by a Restormer backbone optimized via a composite hybrid loss (SSIM, L1, and Sobel gradient) through knowledge distillation, the platform integrates an adaptive high-frequency detail booster alongside multiplicative gain intensity alignment. This architecture delivers 2- to 3-fold objective gains in signal-to-noise ratio (Auto-SNR) and contrast-to-noise ratio (Auto-CNR) while rigorously preserving microstructural integrity and preventing signal dropout holes.Key Technical Features:Multi-Format Ingestion: Seamless batch parsing of volumetric NIfTI formats (.nii.gz) and standard multi-frame clinical DICOM archive packages (.zip) with complete spatial metadata inheritance.Automatic Sequence Routing & Zero-Shot Capability: Automated sequence identification and dynamic weight allocation for PD-FS and T1WI protocols, with verified zero-shot generalizability on non-finetuned T2-FS acquisitions.Non-Destructive Physical Calibration: Robust quantile mapping and morphological mask-guided multiplicative gain alignment, eliminating global baseline shift and negative-clamping boundary artifacts.Integrated ITHP Quality Benchmarking: Built-in Intra-Tissue Homogeneous Patch (ITHP) estimation engine providing slice-level automated quantification of Auto-SNR, Auto-CNR, Diff-PSNR, Removed Noise (%), and Diff-SSIM.Clinical PACS Storage Integration: Native implementation of standard DICOM network services (pynetdicom), enabling C-ECHO ping verification and C-STORE direct transmission to institutional PACS workstations.Cross-Platform Responsive GUI: Lightweight Streamlit-based interface featuring real-time paired slice visualization, multi-language support (English, Chinese, Japanese), and single-click compressed packaging.Note: The uploaded data is a packaged Streamlit program, and none of the uploaded data contains any sensitive information such as patient data. Musculoskeletal MRI; Transformer; Image Denoising; Microstructural Preservation; PACS Integration Hao Wang PUBLIC https://mit-license.org dataset Science Data Bank