Forensic Image Computing: Applications in Bloodstain Pattern Analysis and Virtual Autopsy

Forensic science involves the application of a broad spectrum of scientific principles within a legal framework, and is typically used to support criminal investigations. Drawing from various scientific fields such as physics, chemistry, and biology, forensic science is a vital tool in unearthing the truth in any legal proceeding. Over the course of its existence, forensic science evolved towards an evidence-based framework, relying on rational procedures. Consequently, today's focus of forensic investigators is on the recognition, identification, and analysis of evidence, all made possible through the introduction of countless techniques and technological advances. However, in spite of all technological advances already made, forensic investigations are not solved by the push of a button. Crime laboratories employ large teams of investigators that often have to perform tedious and time-consuming tasks. Such manual work is detrimental to the overall investigation, as it introduces subjectivity into the analysis, which in turn may lead to miscarriages of justice. Moreover, having forensic investigators perform lengthy analyses on a crime scene increases the risk of introducing, removing, or altering crucial evidence. Finally, considering the fact that forensic investigators are highly skilled experts, manual work increases the cost of the overall investigation. In this thesis, our main aim is therefore on replacing time-consuming work by automated procedures in an attempt to tackle the aforementioned problems. More specifically, our focus is on the application of data-driven image computing methods to further advance the automation in the fields of Bloodstain Pattern Analysis and Virtual Autopsy. In the first part of this thesis, we focused on the field of Bloodstain Pattern Analysis. Concretely, we developed a software package called HemoVision, which enables fully automated analysis of impact patterns. In doing so, three major contributions were made. First, we developed a statistical shape model that enables fully automated analysis of individual bloodstains. Compared to existing methods, the proposed approach is faster, more accurate, and more objective. Second, we proposed a vision-based approach using fiducial markers to automatically reconstruct impact patterns from various images. This step was crucial to the development of HemoVision, as it removes most of the manual work that is required with existing methods. Third, we integrated the proposed methods in an intuitive interface, allowing forensic investigators to use our software. Moreover, we performed a validation study to asses HemoVision's overall accuracy. Compared to the state-of-the-art, HemoVision obtains competitive results. In the second part of this thesis, we turned to the field of Virtual Autopsy. More specifically, our focus was on the development of a general framework that enables the automatic detection of abnormalities in medical images. Again, three major contributions were made. First, a data-driven preprocessing method for medical images was developed, allowing to correct imbalances in tissue intensity distribution widths. This is an important step, as standard image processing methods are negatively influenced by this so-called heteroscedasticity. Second, we developed a semi-supervised exemplar-based method to detect abnormalities in medical images. In contrast to existing approaches, the proposed method does not require annotated training data, does not rely on ill-posed non-rigid deformations, and is independent of the imaging modality. Third, we applied the aforementioned methods to the problem of automatic gunshot trajectory reconstruction. Using a robust linear regression model, we were able to correctly estimate trajectories in the majority of the cases. To the best of our knowledge, this is the first method that allows fully automatic reconstruction of gunshot trajectories. Overall, we believe we have contributed significantly to the fields of Bloodstain Pattern Analysis and Virtual Autopsy. By introducing quantitative, automated methods, forensic investigations become less costly, more objective, and faster, which in the end, is beneficial to all people involved in criminal investigations.

Jaar van publicatie:2018

Toegankelijkheid:Open