Jiang, Bohan, McNeil, Andrew J., Liu, Yihao, House, David W., Mbala-Kingebeni, Placide, Mbaya, Olivier Tshiani, Silaphet, Tyra, Dodd, Lori E., Cowen, Edward W., Nussenblatt, Veronique, Bonnett, Tyler, Chen, Ziche, Saknite, Inga, Dawant, Benoit M., & Tkaczyk, Eric R. (2025). Mpox lesion counting with semantic and instance segmentation methods. *Journal of Medical Imaging, 12*(3), 34506. https://doi.org/10.1117/1.JMI.12.3.034506
Mpox is a viral illness with symptoms similar to smallpox, and one important way to track how the disease is progressing is by counting the number of skin lesions (sores) a patient has. Counting these lesions by hand is time-consuming and can easily lead to mistakes. Previously, we created a method to count mpox lesions automatically using a type of computer program called a UNet segmentation model, based on 66 photos from 18 patients. In this study, we tested four additional computer methods—Mask R-CNN, YOLOv8, E2EC, and UNet++—to see how well they could count lesions compared to our original UNet model. We tested each method by leaving out one patient’s data at a time and measuring performance using standard scores that assess accuracy and how close the lesion counts were to the real numbers. The UNet++ model performed best with an F1 score of 0.81, while Mask R-CNN and YOLOv8 both scored 0.75, E2EC scored 0.70, and the original UNet scored 0.79. When we combined all the models into an ensemble, it did not perform better than the best single model (UNet++). This is likely because the different models often made the same mistakes. Overall, both instance segmentation methods and UNet-based methods worked about equally well for counting lesions. The main limit on performance seems to be the number and quality of available photos, not the type of computer method used.
Fig. 3
Lesion-level F1 scores in each photograph relative to the ground truth (human rater). F1 score ranges from 0 to 1, with a higher value indicating a better performance. Each dot indicates each photograph (n=66) of all patients (N=18).
