Deep Learning Methods for Post Mortem Interval Estimation

Large photographic collections of human decomposition with deep learning (DL) techniques suitable for the analysis of human decomposition images would advance research on postmortem interval (PMI) estimation if accurate, high-resolution, and curated datasets and suitable DL methods were created. With this aim, the project will collect donor-specific high-resolution temperature and humidity data, entomological and soil samples, and scavenging events, and integrate the results of these observations with daily photographs taken of each donor at the Anthropological Research Facility of the University of Tennessee, Knoxville. Human decomposition-specific DL techniques will be used to identify the state of the photographic protocol (head-to-shoulders, feet, whole body), to segment (identifying the exact location within each photo) the photos by individual body parts, and to identify distinct stages of decay (via to be developed semi-supervised DL methods). This rich and high-resolution data, together with donors’ demographic information, will be integrated with the ICPUTRD system that allows search, curation, and tagging of over one million photos of human decomposition. The resulting detailed records including images, time since death, and various calculations of Accumulated Degree Days (ADD) proposed in the literature will be used to train specialized image-based DL regression models PMI. The resulting dataset (with images replaced by the DL-produced PMI-salient image-derived feature sets to preserve privacy of donor families) will be published.  Particular attention will be paid to characterizing the PMI-related context variables for the ARF site to both define the context under which the developed PMI predictions are likely to be valid and to determine the types of environmental conditions for which the data like that collected at the ARF is still lacking. We expect the PMI (and ADD) prediction to be directly relevant to casework where the environmental conditions are like those at the ARF site, but the techniques can be expanded in the future to reflect different climatic and environmental conditions. The published dataset of rich and fine-grained observations of factors known to affect PMI with PMI-relevant features extracted from associated photographs and the actual PMI would facilitate the comparison of existing and development of novel PMI methods. The integration of PMI-relevant data into the ICPUTRD system would make its resources, including the ability to search, annotate, and explore a curated collection of over a million of decomposition images, available for study of how various environmental and donor-specific factors affect PMI.