Explainable deep learning in bloodstain pattern analysis: A pilot study using convolutional neural networks with...

The purpose of this pilot study was to explore the feasibility of applying a novel explainable deep learning (XDL) methodology to classify Bloodstain Pattern Analysis (BPA) patterns. A convolutional neural network (CNN) was applied to classify impact and non-impact BPA patterns. A combination of BPA patterns generated by the researcher were supplemented by open-source BPA datasets and used for the CNN training. This methodology yielded promising results of up to 79% accuracy over 10 folds, validating the feasibility of such research avenues in the field. Furthermore, saliency maps were applied as the CNN's explainability layer, which is a novel application of XDL in BPA interpretations. The study highlights the potential of a novel BPA research stream in XAI while also underscoring the potential of the model to act as a possible reliable alternative tool for BPA experts over manual classification methodologies. This approach does not aim to replace the human element from the forensic science process but rather provide a tool to BPA experts to aid and expedite BPA interpretations without compromising on explainability metrics. It has the potential to increase transparency and trust in deep learning systems, which in turn would increase the reliability of forensic outcomes.