基于Deeplabv3+算法的皮肤癣菌菌种分类模型构建与性能验证OA

An artificial intelligence(AI)model based on the Deeplabv3+algorithm was constructed to achieve rapid and accurate classification of clinically common dermatophyte species and enhance the efficiency of traditional diagnosis.In total,74 clinically common dermatophyte strains(including Trichophyton rubrum,T.mentagrophytes,Microsporum canis,and Epidermophyton floccosum)were selected.After verifying the accuracy of the strains by internal transcribed spacer(ITS)sequencing,lactophenol cotton blue staining images were acquired from cultured strains,yielding 962 images.These images were divided into a training set(839 images)and a test set(123 images)for model construction and performance verification,and the results were compared with those of routine visual identification by clinical technicians.The AI model achieved an overall accuracy of 71.54%for the four dermatophyte species,with the per-class recall values being 80.00%for M.canis,79.07%for E.floccosum,65.00%for T.mentagrophytes,and 53.33%for T.rubrum.The overall accuracy of routine visual identification by clinical technologists was 72.36%.Consistency analysis revealed a substantial agreement between the AI model and visual identification(Kappa=0.769,P<0.01).Subgroup analysis showed almost perfect agreement for T.rubrum(Kappa=0.875)and T.mentagrophytes(Kappa=0.808),while the Kappa values for E.floccosum and M.canis were 0.677 and 0.534,respectively.No statistically significant difference was observed in the overall accuracy between the two methods(χ2=0.33,P>0.05).AI model exhibited a significantly higher recall for T.mentagrophytes than visual identification(P<0.05),whereas its identification performance for E.floccosum was significantly lower(P<0.05).In terms of recognition speed,the average speed of the AI model was(2.46±0.04)seconds per image,significantly faster than that of visual identification(3.75±0.61)seconds per image,P<0.05.The AI model could process images in batches continuously(approximately 40 minutes for disposing 962 images)without performance degradation,while visual identification was limited by visual fatigue and required intermittent rest,taking approximately 290 minutes for disposing the same number of images.This is the first construction of an AI classification model for dermatophytes,filling the research gap in this field.The model demonstrated comparable identification accuracy substantially consistent with visual identification of clinical technologists,offering the advantages of faster processing speed,stable batch processing capability,and without restriction of time and location.It can serve as a novel tool for rapid clinical diagnosis,especially in primary medical institutions,and thus holds promising prospects for clinical application and popularization.