Project DetailsAcne Vulgaris is a common skin condition that requires consistent clinical diagnosis, assessment, and follow-up. Disease severity is clinically determined based on a combination of characteristics such as lesion type (e.g., comedones, papules, pustules, and nodules), depth, and the presence of scarring, while lesion counting serves as a central but not exclusive indicator. Current clinical evaluation largely relies on manual lesion counting and subjective severity grading, a process that is time-consuming, dependent on clinician expertise, and prone to inconsistency. This project introduces AcneFusionNet, a deep learning–based system operating under a multi-task learning framework that simultaneously performs acne lesion counting and severity classification from facial images. The main contribution of this work is the development of a dedicated feature fusion mechanism that extracts features from the input image and integrates the predicted lesion count from the detection branch directly into the severity classification head, thereby leveraging the established clinical relationship between lesion quantity and disease severity. In addition, the system employs Label Distribution Learning (LDL) to represent clinical uncertainty in the grading process and utilizes adaptive density maps derived from ground truth annotations for the counting task. The proposed system was evaluated on the ACNE04 dataset and achieved a mean accuracy of 85.86% in severity classification. These results demonstrate the feasibility of the proposed approach as an objective and reliable engineering–medical decision support tool that may support improved clinical assessment and longitudinal monitoring in dermatology, while not replacing comprehensive clinical evaluation.
