Age Prediction of Hematoma from Hyperspectral Images Using Convolutional Neural Networks.

Area of Science:

• Forensic Science
• Biomedical Imaging
• Machine Learning

Background:

• Hematoma age estimation is crucial in forensic science but currently relies on subjective visual analysis.
• Hyperspectral imaging (HSI) offers objective data by capturing detailed spectral information reflecting biochemical changes over time.

Purpose of the Study:

• To evaluate the efficacy of a convolutional neural network (CNN) integrating spectral and spatial data for improved hematoma age estimation.
• To determine if a reduced subset of physiologically relevant wavelengths can maintain estimation accuracy.

Main Methods:

• Hyperspectral imaging (HSI) was used on forearm hematomas from 25 participants.
• Radiometric normalization and SAM-based segmentation were applied to extract hyperspectral patches.
• A CNN model was trained and validated using leave-one-subject-out cross-validation, compared against a spectral-only Lasso baseline.

Main Results:

• The CNN model significantly outperformed the spectral-only baseline, reducing mean absolute error (MAE) from 3.24 to 2.29 days.
• Band-importance analysis identified 20 key wavelengths that, when used alone, matched or exceeded the accuracy of the full 204-band model.
• The optimized approach maintained accuracy across early, middle, and late hematoma stages.

Conclusions:

• Spectral-spatial modeling with CNNs enhances objective hematoma age estimation accuracy.
• Physiologically grounded band selection reduces data dimensionality without compromising performance.
• This approach facilitates the development of practical multispectral systems for clinical and forensic applications.