Histopathological Assessment of Myocardial Ischemia-Reperfusion Injury Using Transformer-Based Artificial Intelligence: Model Comparison Study

Figures

• 

  Figure 1.. Comparative performance of different deep learning models in the recognition of MIRI on H&E-stained histological images. (A) Classification accuracy of 8 representative deep learning models on the test set. The transformer model outperforms all others in overall accuracy. (B) Receiver operating characteristic (ROC) curves for each model. The transformer model achieves the highest AUC value of 0.963. (C) Confusion matrix of the transformer model, demonstrating the lowest false-positive and false-negative rates among all models, indicating superior classification consistency and boundary recognition capability. CNN: convolutional neural network; GAN: generative adversarial network; GCN: graph convolutional network; LSTM: long short-term memory; RNN: recurrent neural network; VAEs: variational autoencoders.

• 

  Figure 2.. Transformers achieve the best overall performance across 5 key metrics. Radar plot showing the aggregated performance of all deep learning models across 5 evaluation metrics—accuracy, precision, recall, F1-score, and area under the curve (AUC). The transformer architecture outperforms all other models across every dimension, demonstrating superior discriminatory power and generalizability, particularly for complex tissue structures and highly heterogeneous lesion patterns. CNN: convolutional neural network; GAN: generative adversarial network; GCN: graph convolutional network; LSTM: long short-term memory; RNN: recurrent neural network; VAEs: variational autoencoders.

• 

  Figure 3.. Training dynamics of the transformer model showing loss reduction and accuracy improvement. (A) Loss function curves for the training and validation sets, showing that the model gradually converged during training without oscillation or overfitting. (B) The accuracy of the training and validation sets increased progressively with the number of iterations, indicating that the model had a stable learning process and good generalization ability.

• 

  Figure 4.. Evaluation of the effects of multiple therapeutic interventions on structural and functional tissue parameters following myocardial ischemia-reperfusion injury. (A-C) Comparisons of necrotic area percentage, apoptotic area percentage, and fibrosis scores among treatment groups indicate that the antioxidant group shows the most significant structural protection. (D-F) Group comparisons of inflammation scores, oxidative stress indices, and capillary density reveal that the antioxidant group performs better in modulating immune responses, providing antioxidant protection, and maintaining microcirculation. Statistical significance was assessed using one-way ANOVA followed by post hoc multiple comparisons. *P<.05; **P<.01; ***P<.001; ns: no significant difference.

• 

  Figure 5.. Evaluation of transformer model performance under 4 therapeutic interventions. (A) The precision-recall (PR) curves of the 4 treatment groups demonstrate the model’s classification accuracy under different intervention conditions, with the antioxidant group showing the highest average precision (AP) value, indicating the strongest sensitivity of the model in this group. (B) The corresponding receiver operating characteristic (ROC) curves indicate that the antioxidant group also has the highest area under the curve (AUC), suggesting optimal model discriminative power for this image category. The calcium channel blocker group shows the weakest performance, implying that the complexity of its image features may pose greater challenges for the model.

• 

  Figure 6.. Heatmap of tissue damage prediction confidence by the transformer model across treatment groups. The heatmap illustrates the distribution of damage prediction confidence scores for 20 myocardial hematoxylin and eosin–stained image samples per group under 4 treatment conditions. The color gradient ranges from blue (low confidence) to red (high confidence), indicating the model’s predicted probability of tissue damage. Samples from the control group are predominantly concentrated in high confidence regions (red), while those from the antioxidant group are primarily located in low-confidence areas (blue), reflecting the model’s discriminative sensitivity and stratification stability across different treatment backgrounds.

• 

  Figure 7.. Dynamic trends of key myocardial injury indicators at 3 time points across different treatment groups. (A) The necrotic area ratio over time shows the most substantial reduction in the antioxidant group by 72 h, indicating a strong tissue-protective effect. In contrast, the control group exhibits a continuous upward trend, suggesting progressive damage without intervention. (B) A similar trend is observed for the apoptotic area ratio, with the antioxidant group showing the most significant decline, while other groups demonstrate only mild decreases, and the control group slightly increases. Error bars represent each group’s SD.

• 

  Figure 8.. Evaluation of classification performance of the transformer model on myocardial injury images at 3 time points. (A) The receiver operating characteristic (ROC) curve comparison indicates that the model achieves the highest area under the curve (AUC) of 0.98 at 72 h, demonstrating excellent sensitivity to mild late-stage injury. The AUC values at 24 and 6 h are 0.97 and 0.91, respectively, confirming the model’s stable performance in identifying acute-phase lesions. (B) The precision-recall (PR) curve demonstrates that the model achieves the highest balance between precision and recall at 72 h, confirming its generalization capability across time and consistent performance in early- and late-stage recognition.

• 

  Figure 9.. Workflow and performance of the MIL-based weakly supervised framework. (A) Schematic diagram of the weakly supervised learning pipeline. Whole-slide images (WSIs) are partitioned into 512×512-pixel tiles and assigned slide-level labels based on treatment group and reperfusion time. Using the CLAM framework, the model applies an attention mechanism to automatically generate pseudo-regions of interest (ROIs) and learn pathological discriminative features. (B) Representative H&E image overlaid with the attention heatmap. Red regions denote high-importance areas identified by the model. (C) Receiver operating characteristic (ROC) curve of the weakly supervised transformer classifier, showing an area under the curve (AUC) of 0.947. (D) Precision-recall (PR) curve illustrating the balance between precision and recall across varying thresholds. The attention-derived pseudo-ROIs generated from this workflow are subsequently used to guide model pretraining.

• 

  Figure 10.. Consistency analysis between model outputs and pathologist measurements for necrotic and apoptotic regions. Top row: Pearson correlation analyses comparing model-predicted necrotic and apoptotic areas with manual measurements by expert pathologists. Strong linear positive correlations were observed for both necrosis (r=0.93) and apoptosis (r=0.89), with all P<.001. The dashed line represents the line of perfect agreement. Bottom row: Bland-Altman plots assessing agreement between model predictions and expert annotations. The mean differences were 0.002 for necrotic regions and 0.003 for apoptotic regions, with 95% limits of agreement within +5% and −5% and no systematic bias detected. Blue and orange points represent necrotic and apoptotic samples, respectively. These results demonstrate a high degree of concordance between the transformer model and expert assessments, confirming the model’s reliability and clinical consistency in identifying myocardial injury regions.

• 

  Figure 11.. Linear correlation between model predictions and expert pathological scores. The scatter plot illustrates the correspondence between transformer model predictions and expert manual assessments. The red line represents the linear regression trend, with the shaded area indicating the 95% CI. The Pearson correlation coefficient is 0.886 with a P value of 5.19×10-21, indicating a strong positive correlation and validating the model’s consistency and reliability in assessing tissue injury severity.