Classification of wound types with deep learning models using forensic data

In this research, we present DRDarkNet, a hybrid deep feature engineering model designed for the automatic classification of cause of death using autopsy images. The primary objective is to demonstrate the model's classification capability and its efficacy in accurately categorizing autopsy images into six classes: crush, choking, cutter-punisher, firearms, burn, and drowning. The proposed architecture comprises four phases: deep feature extraction, multiple selectors-based feature selection, classification, and information fusion. We collected a dataset containing 4254 autopsy images for training and evaluation purposes. The deep feature extraction phase utilizes pretrained DenseNet201, ResNet50 and DarkNet53 models to extract six feature vectors from the convolutional neural networks (CNNs). These vectors are then used to construct DRDarkNet . To identify the most informative features, we employ three feature selectors, namely Chi2, ReliefF, and neighborhood component analysis (NCA), resulting in 18 selected feature vectors. Support vector machine (SVM) classifiers are applied to these selected vectors, generating 18 classifier-wise results. The information fusion phase incorporates a pruning-based iterative majority voting (PIMV) technique to compute the voted results and select the optimal output as the final classification result. DRDarkNet outputs both classifier-wise results and voted results, achieving a final classification accuracy of 96.47%. The accuracy of over 96% across all six classes demonstrates the effectiveness of our proposed model for autopsy image classification. The successful classification results highlight the efficacy of DRDarkNet as a reliable model for autopsy image analysis. The high accuracy in identifying cause of death showcases the potential of deep feature engineering techniques in the field of forensic pathology. DRDarkNet holds promise as a valuable tool to assist forensic experts and contribute to the advancement of autopsy investigations.