Collaborative Artificial Intelligence Platform for Bruise Age Analysis

Law enforcement and prosecutors rely heavily on the accuracy and interpretation of injury documentation to inform their decisions. However, bruise assessments by forensic clinicians provide little valid or reliable data about its age beyond qualitative, subjective opinion. Image analysis using deep learning (DL), a sub-domain of machine learning, has demonstrated significant benefits in accuracy and reliability within healthcare; yet few studies have applied the techniques to forensic trauma analysis. Purpose: The purpose of our project is to: 1) determine whether advanced, time-series approaches to DL models of images can improve our understanding of how bruise appearance changes over time on diverse skin tones; and 2) develop a bruise image platform with DL modeling to support future research and collaboration in forensic science. Methods: Phase 1 (Years 1-2) will focus on investigating potential pathways for leveraging DL in the analysis of bruise imaging. A large (~26,000) dataset of bruise images of known age and diverse skin colors will be used to prototype and analyze advanced DL models (e.g., recurrent neural networks, attention mechanisms, manifold learning). Metrics will include predictive accuracy, false positive and negative rates, Receiver-Operator-Characteristic curves, and confusion matrices. An image collaboration platform will also be developed during Phase 1 based on likely case uses by forensic clinicians and researchers. The platform’s architecture will follow a standard design with web-based user interface. Image metadata and structured data will be stored in a custom-designed relational database that can be deployed on a dedicated system in the cloud. The platform will ultimately be populated with the dataset, including both unprocessed and processed files. In Phase 2 (Year 3) the prototyped and evaluated neural networks will be integrated into the image collaboration platform. A mixed-methods study will be conducted with forensic clinicians and researchers to compare the model’s accuracy to expert opinion and to investigate the feasibility of the platform for end-user collaboration. Deliverables: Deliverables will include submission of peer-reviewed publications, conference presentations, and a webinar. The integrated image collaboration platform will be developed to be publicly accessible, demonstrable, and scalable as well as sustainable with future funding. Implications: By leveraging our university’s resources in DL and a vast image library, we will develop new, quantitative approaches to the forensic analysis of bruises while establishing limits of reliability
52 and accuracy for different injury conditions. The results will have broad implications for forensic clinical practice, criminal justice response, and future research.