Deep learning for automated boundary detection and segmentation in organ donation photography

Area of Science:

• Computer Vision
• Medical Imaging
• Surgical Technology

Background:

• Medical photography is crucial in surgery.
• Accurate image segmentation is vital for computer vision analysis.
• Automated segmentation for kidney and liver donation images is underdeveloped.

Purpose of the Study:

• To develop and compare deep learning segmentation models for kidney and liver organ donation photographs.
• To evaluate the performance of novel models against existing background removal tools.

Main Methods:

• Developed two deep learning models: Detectron2 and YoloV8, using transfer learning.
• Utilized anonymized datasets of 821 kidney and 400 liver images for training/internal validation, and 203 kidney and 208 liver images for external validation.
• Assessed segmentation performance using Intersection over Union (IoU) for whole organ and clear view (parenchyma only) labels.

Main Results:

• Detectron2 and YoloV8 achieved high IoU scores (0.92-0.97) in both whole kidney and liver segmentation, outperforming other models (max IoU 0.59).
• Segmentation tasks were completed rapidly, within 0.13-1.5 seconds per image.
• Both models demonstrated superior performance on both internal and external validation datasets.

Conclusions:

• Open-source deep learning software enables accurate, rapid, and automated image segmentation for surgical photography.
• These advanced models significantly outperform existing methods.
• The findings could drive advancements in surgical computer vision, similar to progress in other medical imaging fields.