Automatic detection and identification of diatoms in complex background for suspected drowning cases through object...

Area of Science:

• Forensic Medicine
• Algology
• Computational Biology

Background:

• Diagnosing drowning is challenging in forensic medicine.
• The diatom test aids drowning diagnosis by detecting diatoms (unicellular algae) in samples.
• Current methods using enzyme digestion for diatom observation create debris, hindering accurate identification.

Purpose of the Study:

• To develop a novel AI model for accurate diatom detection and identification in forensic samples.
• To address limitations of existing methods in complex backgrounds under light microscopy.
• To improve the accuracy and efficiency of the diatom test in drowning investigations.

Main Methods:

• Development of a novel model using sequential transfer learning of object detection models.
• Training and validation of models on forensically relevant diatom species.
• Comparison of sequential transfer learning models against traditional transfer learning approaches.

Main Results:

• The developed models accurately detect and identify up to 50 diatom species.
• Average precision and recall range from 0.7 to 1, varying by species.
• Models developed via sequential transfer learning outperformed those from traditional transfer learning.

Conclusions:

• A novel AI-based method using sequential transfer learning shows high accuracy for diatom detection and identification.
• This approach offers a significant improvement over existing methods for forensic diatom analysis.
• The best performing model is anticipated for routine use at the Medicolegal Institute of Paris.