Age Prediction of Hematoma from Hyperspectral Images Using Convolutional Neural Networks

Accurate estimation of hematoma age remains a major challenge in forensic practice, where assessments rely largely on subjective visual interpretation. Hyperspectral imaging (HSI) captures rich spectral signatures that may reflect the biochemical evolution of hematomas. We evaluate whether integrating spectral and spatial information with a convolutional neural network (CNN) improves hematoma age estimation and whether a reduced, physiologically motivated subset of wavelengths can maintain performance. A forearm hematoma dataset from 25 participants was processed using radiometric normalization, SAM-based segmentation, and extraction of 64 × 64 × 204 hyperspectral patches. Using leave-one-subject-out cross-validation, the CNN achieved substantially lower mean absolute error (MAE 2.29 days) compared to a spectral-only Lasso baseline (MAE 3.24 days). Bandimportance analysis combining SmoothGrad and occlusion sensitivity identified 20 highly informative wavelengths, and using only these bands matched or exceeded the accuracy of the full 204-band model across early, middle, and late hematoma stages. These results show that spectral-spatial modeling and physiologically grounded band selection can significantly enhance hematoma age estimation while reducing data dimensionality, supporting the development of compact multispectral systems for objective clinical and forensic evaluation.