Mpox lesion counting with semantic and instance segmentation methods.

Area of Science:

• Medical imaging
• Computer vision
• Infectious disease modeling

Background:

• Mpox (mpox) is a viral illness presenting symptoms similar to smallpox.
• Accurate monitoring of mpox progression relies on quantifying skin lesions.
• Manual lesion counting is time-consuming and prone to human error.

Purpose of the Study:

• To compare the performance of various AI models for automated mpox lesion counting.
• To evaluate instance segmentation (Mask R-CNN, YOLOv8, E2EC) and semantic segmentation (UNet, UNet++) methods.
• To determine if an ensemble of models improves lesion counting accuracy.

Main Methods:

• Four AI models (Mask R-CNN, YOLOv8, E2EC, UNet++) were compared against a baseline UNet model.
• A patient-level leave-one-out cross-validation strategy was employed.
• Performance was assessed using F1 score and Bland-Altman analysis for lesion count agreement.

Main Results:

• UNet++ achieved the highest F1 score (0.81), closely followed by the baseline UNet (0.79).
• Mask R-CNN and YOLOv8 achieved F1 scores of 0.75, while E2EC scored 0.70.
• Bland-Altman analysis revealed the narrowest limits of agreement for UNet++ (62.1) and UNet (69.1).

Conclusions:

• Both instance and semantic segmentation models demonstrate comparable efficacy in mpox lesion counting.
• An ensemble of models did not outperform the best single model (UNet++), suggesting shared error patterns.
• The study highlights that data quality and quantity, rather than algorithmic choice, may be the primary limitation for improving performance.