锐器创自动识别与分类OA

Objective To evaluate the feasibility of automatic identification and classification of sharp wounds using deep learning network models.Methods A total of 1 475 images of stab wounds,chop wounds,slash wounds,and shear wounds were collected and divided into training,validation,and test sets at an 8∶1∶1 ratio.After preprocessing,the images were input into fine-tuned models based on three pre-trained classification networks:Vit-L32-21k,Densenet-201,and Efficientnet.The model was evaluated using precision,accuracy,recall,F1 score,human-machine confrontation analysis and reading time as metrics,with results visualized through heat maps.Results The model achieved an overall classification accuracy and recall of 75.0%-81.6%,with an F1 score above 0.749 and reading time(<0.1 s)significantly shorter than that of forensic pathologists.Among the four sharp wounds,stab wounds(96.4%)and chop wounds(77.5%)achieved classification accuracy comparable to that of senior forensic pathologists,while shear wounds(60.0%)and slash wounds(47.3%)showed lower accuracy,comparable to that of junior forensic pathologists.Classification accuracy was positively correlated with sample size.Heat maps revealed trauma features consistent with what was observed by forensic pathologists during classification.Conclusion The model demonstrated the ability to automatically identify and classify stab and chop wounds with accuracy comparable to that of senior forensic pathologists,thus providing visualized classification rationales through heat maps.