AUTHOR=Zhang Daojun , Li Huanyu , Shi Jiajia , Shen Yue , Zhu Ling , Chen Nianze , Wei Zikun , Lv Junwei , Chen Yu , Hao Fei 

TITLE=Advancements in acne detection: application of the CenterNet network in smart dermatology

JOURNAL=Frontiers in Medicine

VOLUME=Volume 11 - 2024

YEAR=2024

URL=https://www.frontiersin.org/journals/medicine/articles/10.3389/fmed.2024.1344314

DOI=10.3389/fmed.2024.1344314

ISSN=2296-858X

ABSTRACT=In the field of acne detection, the quality control of acne imagery, coupled with precise segmentation and grading, holds paramount importance. However, current research has mostly focused on a single aspect of these elements. In contrast, our proposed multi-task acne detection method aims to comprehensively address these critical issues. We employed a CenterNet-based training paradigm to architect an advanced acne detection system. This model, which collects acne images through smartphones, possesses the multi-tasking capability of simultaneously detecting image quality and various acne types.Within its diagnostic schema, the system not only effectuates the differentiation of lesions such as comedones, papules, pustules, and nodules but also integrates a meticulous delineation capability for cysts and post-acne scars. This facilitates a more holistic and precise lesion identification and enumeration. The efficacy of this approach was subsequently corroborated in clinical diagnostic environments. The results show that our sophisticated acne detection system boasts a lesion categorization accuracy reaching 83%, marking a 12% superiority over ResNet18 models. Furthermore, our model manifested a stratification precision of an 76%, surpassing dermatologists by 16%.In this study, we introduce a multi-task learning-based acne detection framework that seamlessly integrates classification, localization, counting, and precise segmentation of acne. This model not only allows dermatologists to identify acne lesions more accurately during remote diagnosis, but also helps them clearly understand the grading logic and criteria and make grading judgments easily.