基于可解释性的Transformer模型在早期尘肺病辅助诊断中的临床价值评估OA

[Background]The diagnosis of early-stage pneumoconiosis relies heavily on physician experi-ence,leading to substantial subjectivity and interobserver variability.The clinical value and inter-pretability of the Vision Transformer(ViT)architecture for the early identification of pneumoco-niosis require systematic validation. [Objective]To systematically evaluate the diagnostic performance of ViT-based deep learning models for identifying early-stage pneumoconiosis on digital radiography(DR)chest radiographs,and to enhance the transparency of model decision-making process using the gradient-weighted class activation mapping(Grad-CAM)algorithm,thereby providing a reliable basis for developing computer-aided diagnosis tools. [Methods]This study retrospectively collected high-kilovolt chest radiographs from 1 326 dust-exposed workers.The lung fields were au-tomatically segmented using a U-Net model trained with a hybrid loss function combining Dice loss and cross-entropy loss,together with data augmentation techniques.Subsequently,using the segmented lung field regions of interest(ROIs),four ViT-derived models(ViT,SimpleViT,TwinsSVT,and CrossFormer)were trained,and their diagnostic efficacies for the binary classification(normal vs.abnormal)of early-stage pneumoconiosis were compared.The consensus diagnosis of three senior radiologists served as the reference standard.The dataset was randomly divided into a training set(928 cases)and a validation set(398 cases)at a 7∶3 ratio.All classification models were trained using the Adam optimizer for 3 000 epochs with an initial learning rate of 0.001,and a cosine decay strategy was employed to dy-namically adjust the learning rate.Model performance was compared using DeLong's test,diagnostic consistency was assessed via the Kappa coefficient,and the optimal model was further evaluated by feature visualization analysis using Grad-CAM. [Results]The U-Net lung field segmentation model demonstrated excellent performance on the validation set,achieving a Dice coefficient of 0.971 and a mean intersection over union(mIOU)of 0.944.In the classification task,the CrossFormer model performed best,yielding an area under the curve(AUC)of 0.962(95%CI:0.945,0.979),with accuracy,sensitivity,specificity,and F1-score of 92.0%,92.1%,91.8%,and 0.916,respectively.Its diagnostic performance was superior to the other evaluated models(DeLong's test,P<0.01).The diagnostic performance of the CrossFormer model was highly comparable to that of senior radiologists(Kappa=0.834,P<0.001),with its accuracy(91.7%)being on par with Physician A(92.5%)and Physician B(92.2%).McNemar's test indicated no systematic bias in diagnostic errors(P>0.05).Grad-CAM visualization revealed that the model's attention maps aligned closely with the typical radiological distribution of pneumoconiosis,mainly involving the upper and middle lung zones and perihilar regions,corresponding to the actual distribution of small opacities. [Conclusion]The CrossFormer model demonstrates excellent diagnostic performance for the early identification of pneumoconiosis.Combined with Grad-CAM-based interpretability analysis,the model shows clinically plausible decision-making patterns,providing a reliable model selection basis and technical reference for the development of computer-assisted diagnostic tools for pneumoconiosis.