Digital whole-slide image analysis for automated diatom test in forensic cases of drowning using a convolutional neural network algorithm

Abstract

Diatom examinations have been widely used to perform drowning diagnosis in forensic practice. However, current methods for recognizing diatoms, which use light or electron microscopy, are time-consuming and laborious and often result in false positive or negative decisions. In this study, we demonstrated an artificial intelligence (AI)-based system to automatically identify diatoms in conjunction with a classical chemical digestion approach. By employing transfer learning and data augmentation methods, we trained convolutional neural network (CNN) models on thousands or tens of thousands of tiles from digital whole-slide images of diatom smears. The results showed that the trained model identified the regions containing diatoms in the tiles. In an independent test, where the slide samples were collected in forensic casework, the best CNN model demonstrated a performance competitive with those of 5 forensic pathologists with experience in diatom quantification. This pilot study paves the way for future intelligent diatom examinations; many efficient diatom extraction methods could be incorporated into our automated system.