At least 5 million acts of domestic violence occur annually to women aged 18 years and older, according to the CDC. Some victims report this violence to police, which launches a forensic investigation. In abuse cases, evidence to support a victim’s complaints is critical, and bruise dating is one way to establish a timeline in the legal case.
Coroners are typically the ones to date a bruise using a color scale. However, it is a complicated, subjective process that is rife with inaccuracies—published accuracy rates in previous studies are only between 40 and 50%. This is due to a number of factors, including bruise location, size, depth, degree and especially race of the victim. Previous bruising studies have included mainly white-skinned people, opening up a huge research gap for those with darker skin.
Now, researchers in Peru are proposing a model based on deep convolutional neural networks as an alternative for bruise dating—one that is almost twice as accurate as current methods.
In this study, recently published in the Journal of Forensic Science, the researchers tested four different neural networks to see which would provide the most accurate results. But first, they had to generate a dataset robust enough to create classifications of bruises based on age, while also training the networks to interpret those classifications. To do this, they carried out a controlled experiment using a common bruise generation method.
After participating in a paintball match, 11 volunteers of mixed skin took five daily photographs of bruises sustained from game day (day 1) to day 30. Overall, a total of 2,140 photographs were collected, of which 2021 contained a bruise and 119 depicted healthy skin.
“This way, over a period of 60 days, bruise photographs were collected and a dataset was built with detailed characteristics, which includes the number of images that were used for training, validation, and testing of the model, for each of the classes used by forensic doctors in Peru,” the paper explains.
The photographs, captured according to the researchers’ given protocol, are then preprocessed to obtain a clean and segmented image of the bruise. The processed images were then fed into the four neural networks for training. Models based on InceptionV3, Resnet50 and MobileNet showed dating precision less than 57%. However, MnasNet‐based models demonstrated precision greater than 95%, the best results being 97% precision, 97% sensitivity and 99.5% specificity.
“These results confirm the proposed model is suitable for bruise dating, given that it has presented 97% precision for bruise dating on people of mestizo complexion, far above the 50% precision obtained by experts through images of white people, suggesting the model could be used for other skin colors,” the authors conclude.
While processing the photographs, the researchers noted that bruises on the two volunteers with darkest skin were visible up until the fifth day, while bruises on the two people with lightest skin were still visible on day 30. Extending the model to include biological variability, such as age, sex, skin color, and race of the subject, is one area the researchers said should be explored in future work. Taking into account biological status, such as those with chronic diseases that cause bruising more easily, is also an area that needs further examination.
Since the neural network is intended to stand up in court during abuse cases, the researchers said they would also like to extend the model to better incorporate unique aspects of physical violence, such as commonly used objects, intensity and geographical location.
Photo: Preprocessed image of an experimental bruise. Credit: Tirado J, Mauricio D. Bruise dating using deep learning. J Forensic Sci. 2020. https://doi.org/10.1111/1556-4029.14578