Multiparametric Deep Learning To Improve Cardiac Sarcoidosis Detection from Cardiovascular Magnetic Resonance Images

Background: Cardiac sarcoidosis (CS) is challenging to diagnose due to its heterogeneous presentations. Cardiovascular magnetic resonance (CMR) with late gadolinium enhancement (LGE) is the reference standard for detecting myocardial damage, offering diagnostic and prognostic value, but its patchy, variable patterns often overlap with other cardiomyopathies, limiting specificity. Cine CMR provides complementary functional insights, such as wall motion and ventricular performance. Despite the promise of deep learning (DL) for disease detection, most existing methods rely on unimodal inputs and fail to capture the full complexity of CS[1]. We propose a multiparametric DL framework integrating CINE and LGE to capture complementary structural and functional features and improve the accuracy of CS detection.

Methods: We used LGE and CINE CMR images from two cohorts: patients with CS (90) and all-comers without CS (77). Endo and epicardial boundaries for both CMR types are semi-automatically labeled using suiteHEART. LGE and CINE slices are cropped to the left ventricular (LV) myocardium. Post pre-processing, LGE images are interpolated trilinearly and CINE images are uniformly sampled to ensure uniform number of spatial slices. CINE-CMR has 25 timeframes per subject. A pretrained LGE segmentation model[2,3] is used to predict scar regions, used as additional input for classifier training. Cohorts were split into 8:1:1 for train, test and validation sets. We built a triple-encoder framework with ResNet[4] encoders for LGE, predicted scar and CINE inputs. Encoded features are concatenated channel-wise and passed through fully connected layers to generate class probabilities. The model is trained using Cross Entropy[5]. Fig.1 shows the classification framework. To further assess the contribution of CINE, we also evaluate classifier performance using LGE and predicted scar alone.

Results: Fig.2 summarizes classification performance across five independent runs, reporting accuracy, precision, recall, and F1 score for multiparametric vs. single CMR. The multiparametric input achieved higher accuracy of 76% compared to 69% for the single-CMR model. Precision, recall, and F1 score were also higher for the combined method. Fig.3 shows the ROC curves, with the multiparametric approach achieving higher AUC, demonstrating that integrating multiple CMR images with predicted scar regions improves classification performance.

Conclusion: Our framework effectively combines CINE-CMR and LGE-CMR with automatically predicted scar regions to distinguish between subjects with cardiac sarcoidosis. By using two CMR types as inputs with additional information on scar regions, the model captures more comprehensive information, thus resulting in higher classification performance. Future work will involve incorporating more cases of CS from multi-center datasets, using additional tissue characteristics such as T1/T2 maps and benchmarking against expert clinical diagnosis.

Figure 1: Our framework for (i) pre-processing the LGE slices and their annotations using LV myocardial boundaries, where slices are cropped, masked to the left-ventricular myocardium (ROI), resized to 128x128 via bilinear interpolation, and stacked via trilinear interpolation (ii) using the pre-trained segmentation network to obtain the predicted scar regions for each slice, and (iii) training a triple-encoder ResNet framework to predict the probability of cardiac sarcoidosis for the given subject.
Figure 2: Quantitative metrics to validate the performance of our classification framework showing the accuracy, precision, recall and F1 score across all 5 folds of cross validation for LGE + Predicted Scar vs. CINE + LGE + Predicted Scar.
Figure 3: Receiver Operating Characteristic (ROC) curves and corresponding Area Under the Curve (AUC) values to evaluate the performance of our classification framework across all 5 folds of cross-validation for LGE + Predicted Scar vs. CINE + LGE + Predicted Scar. The consistently higher AUC values across folds indicate superior and stable discriminative performance of the combined CINE + LGE + Predicted Scar model compared to LGE + Predicted Scar alone.