Non-Contact Assessment of Diabetic Foot Healing Status Based on Hyperspectral Reconstruction

Diabetic foot, as a common complication of diabetes, is characterized by high recurrence rates, prolonged treatment cycles, and substantial treatment costs. Timely and objective assessment of the healing status of diabetic foot ulcers is of great significance for clinical decision-making. However, existing evaluation methods often rely on physicians' experience and lack unified objective indicators. Hyperspectral imaging can simultaneously acquire spatial and spectral information of tissues, thereby extracting physiological parameters including tissue oxygen saturation (S_SO2), which is closely associated with microcirculatory function and tissue healing potential, and can be used to assess the healing status of foot ulcers in diabetic foot patients. Conventional hyperspectral imaging systems are typically expensive and require specialized equipment, which limits their widespread clinical application. This paper proposes a non-contact diabetic foot healing assessment method based on reconstructed hyperspectral data, enabling quantitative analysis of the local microcirculatory state and effective prediction of healing trends. This method offers the advantages of being non-invasive, low-cost, and highly scalable, providing a new technical approach for early screening of diabetic foot and personalized treatment decisions. This paper proposes a non-contact diabetic foot healing assessment method based on hyperspectral reconstruction from visible light images. First, the HardNet-DFUS segmentation network is employed to perform precise segmentation on visible light images of diabetic foot ulcers (DFUs), obtaining masks of the ulcer region and the foot region, thereby effectively eliminating background interference and localizing the regions of interest. Subsequently, the multi-stage spectral-wise Transformer (MST++) network is utilized to perform hyperspectral reconstruction on the segmented visible light images of the foot region, recovering the reflectance spectral information across 31 channels within the 450–600 nm wavelength range. Based on the modified Beer-Lambert law, S_SO2 is estimated from the reconstructed spectra, and a calibration-free differential S_SO2 feature extraction method is further proposed. This method characterizes the local microcirculatory status by calculating the S_SO2 difference between the ulcer edge region and the distal foot region, effectively avoiding interference from factors such as eschar, necrotic tissue, and illumination variations with the measurement results. Finally, a joint prediction model for diabetic foot healing status based on a support vector machine (SVM) is constructed by integrating the statistically significant clinical physiological parameter, low-density lipoprotein cholesterol (LDL-C), with the differential S_SO2 feature. Experimental results demonstrate that this model exhibits high accuracy and stability on a small-sample dataset of 46 cases, providing an objective and reliable method for assessing diabetic foot ulcer healing. The MST++ network achieved satisfactory hyperspectral reconstruction performance, with a peak signal-to-noise ratio (PSNR) of 11.652, a mean relative absolute error (MRAE) of 0.008, and a root mean square error (RMSE) of 0.0126. The reconstructed spectra exhibited generally consistent spectral trends with the original data, although slightly higher reflectance values were observed beyond 460 nm, primarily attributed to the underestimation of original measurements caused by insufficient illumination during acquisition. The S_SO2 distribution maps generated from the reconstructed data were consistent with the original results in terms of overall trends, indicating that the proposed method effectively preserved the functional information. The SVM model based on reconstructed S_SO2 features (SVM-Re-S_SO2(mine)) achieved a precision of 76.1%, an accuracy of 76.0%, a recall of 68.8%, and an area under the curve (AUC) of 0.667 for predicting ulcer healing, showing slightly lower performance compared to the model using original hyperspectral data. Through logistic regression analysis, LDL-C was identified as the sole clinical parameter significantly associated with healing outcomes (p<0.05). After integrating LDL-C with the reconstructed S_SO2 features, the model performance was substantially improved, achieving a precision of 81.4%, an accuracy of 82.0%, a recall of 76.5%, and an AUC of 0.872, representing increases of 19.7%, 7.9%, 18.2%, and 30.7%, respectively, compared to the S_SO2-only model. This result approached the performance achieved by combining original hyperspectral data with LDL-C, demonstrating that multimodal feature fusion can effectively compensate for reconstruction errors and significantly enhance the discrimination capability for non-healing cases. This study validates the effectiveness of the proposed method for non-contact assessment of diabetic foot healing. This paper proposes a novel method for assessing diabetic foot healing by combining hyperspectral reconstruction with clinical physiological parameters. We collected visible light images, hyperspectral images, and corresponding clinical data from patients, and developed a local tissue oxygen extraction method using the distal foot as a reference. Specifically, the local microcirculatory status is characterized by calculating the S_SO2 difference between the ulcer edge and the distal region. Compared with traditional approaches that focus solely on the ulcer area, this differential feature effectively avoids interference from eschar and necrotic tissue, providing a more accurate reflection of local oxygenation levels. Based on this, we identified LDL-C as a clinical parameter significantly associated with healing outcomes and integrated it with the S_SO2 feature to construct an SVM prediction model. The results show that the fusion model significantly outperforms the single-feature model, with the AUC increasing from 0.667 to 0.872 and accuracy reaching 82.0%. This demonstrates that multimodal feature fusion can effectively enhance predictive capability and validates the feasibility of reconstructing hyperspectral information from visible light images. Overall, the proposed method is non-contact, low-cost, and requires no complex calibration, showing promising potential for clinical application in early screening of diabetic foot and providing objective support for treatment decisions.