Ilmu Komputer & AI editorial
Open AccessOA2026
PP-Net: A Hybrid Physical-Prior Neural Network for Scattered Light Removal in Biomedical Images on Embedded Devices
A three-stage deep learning framework combining denoising, physics-based scattering estimation, and guided refinement for real-time IoMT deployment
Yongfei Guo; Tingjin Chu; Mengzhuo Liu; Hongwei Lou; Yuanhao Gong· 2026· DOI 10.48550/arXiv.2609.26474
The core problem
Scattered light is a pervasive degradation in biomedical imaging, particularly in microscopic and endoscopic scenarios where tissues and fluids scatter incident light. Removing this scattering is challenging due to three factors: (1) aligned scattered-light-free ground truth is rarely available for training; (2) scattering often co-occurs with weak illumination and sensor-induced noise; and (3) many learning-based restoration models are too computationally heavy for embedded devices in Internet of Medical Things (IoMT) settings. This paper introduces PP-Net, a hybrid physical-prior neural network designed to address these issues. The method comprises three components: DFN-Net for sensor noise suppression, ASAP for scattering map estimation and physics-based prior recovery, and GF-Net for refining the prior map by fusing it with the denoised observation. A progressive synthetic training and cross-domain transfer strategy reduces reliance on paired biomedical ground truth. Experiments demonstrate significant improvements in PSNR and SSIM under joint noise-and-scattering degradation, and real-world validation on W2S biomedical images shows a 43.3% reduction in NIQE. Edge deployment vi
Innovation
Quantitative evaluations demonstrate the effectiveness of PP-Net. On paired synthetic benchmarks, the physical-prior branch improves PSNR by up to 1.26 dB. Under joint noise-and-scattering degradation, PP-Net outperforms representative baseline methods by more than 10.8 dB in PSNR and more than 0.62 in SSIM. On real W2S biomedical images, the proposed method reduces the average NIQE score by 43.3%, indicating significant perceptual quality improvement. For edge deployment, RKNN conversion and INT8 quantization achieve an average inference latency of approximately 200 ms per image across 360 test images. These results confirm that PP-Net provides both high restoration quality and practical deployability for embedded IoMT devices.
Scattered light is a pervasive degradation in biomedical imaging, particularly in microscopic and endoscopic scenarios where tissues and fluids scatter incident light. Removing this scattering is challenging due to three factors: (1) aligned scattered-light-free ground truth is rarely available for training; (2) scattering often co-occurs with weak illumination and sensor-induced noise; and (3) many learning-based restoration models are too computationally heavy for embedded devices in Internet of Medical Things (IoMT) settings. This paper introduces PP-Net, a hybrid physical-prior neural network designed to address these issues. The method comprises three components: DFN-Net for sensor noise suppression, ASAP for scattering map estimation and physics-based prior recovery, and GF-Net for refining the prior map by fusing it with the denoised observation. A progressive synthetic training and cross-domain transfer strategy reduces reliance on paired biomedical ground truth. Experiments demonstrate significant improvements in PSNR and SSIM under joint noise-and-scattering degradation, and real-world validation on W2S biomedical images shows a 43.3% reduction in NIQE. Edge deployment via RKNN conversion and INT8 quantization achieves ~200 ms inference per image over 360 test images, making PP-Net suitable for microscopic imaging, endoscopic inspection, and edge-assisted biomedical analysis in IoMT scenarios.
PP-Net is a three-stage pipeline that integrates physical priors with deep learning. The architecture is illustrated below:
Why it matters
The success of PP-Net stems from its hybrid design that combines physical priors with data-driven learning. The three-stage architecture effectively decouples noise suppression, scattering estimation, and refinement, allowing each component to specialize. The progressive synthetic training and cross-domain transfer strategy mitigates the lack of paired ground truth, a common bottleneck in biomedical imaging. The substantial PSNR and SSIM gains under joint degradation highlight the importance of explicitly modeling scattering physics. The 43.3% NIQE reduction on real images validates the method's generalization to practical scenarios. The ~200 ms latency on embedded devices demonstrates that PP-Net is not only accurate but also efficient, making it suitable for real-time applications in microscopic imaging, endoscopic inspection, and edge-assisted biomedical analysis. Future work may extend PP-Net to video sequences and other modalities, and further optimize the model for even lower-power devices.
Who should read this
CS practitioners and researchers
Opening member content…