DeepOx: A Model-Free Reconstruction of O2-Saturation Maps from Cardiac MRI

Background: Prior T2 MRI methods to estimate intracardiac blood oxygen saturation (O₂Sat) rely on explicit models [1-4] that make simplifying assumptions, leading to measurement bias and limited generalizability. We propose DeepOx, a deep-learning-based method that predicts biventricular O₂Sat maps directly from a series of T2-prepared images and does not require additional patient-specific inputs. We aimed to optimize DeepOx in terms of the loss function and training strategy and evaluate against previously shown model-based approach [3].

Methods: A series of six single-shot T2-prepared bSSFP short-axis images in 32 patients at 1.5T (38 datasets total) previously recruited from a single-center for model-based MR oximetry were utilized [3]. A U-Net-based convolutional neural network (CNN) was trained to predict pixel-wise O₂Sat maps (Figure 1). Right heart catheter and pulse oximeter-based reference O₂Sat values were uniformly assigned to pixels of the right and left ventricles (RV and LV), respectively, to generate ground truth O₂Sat maps for training. To enhance generalizability and to encourage spatially invariant learning, we applied extensive data augmentations, including flipping, rotation, translation, and random pixel masking to the images. An ablation study was conducted to evaluate different loss functions - mean squared error (MSE), mean absolute error (MAE), and Huber loss to DeepOx. Using the best-performing loss function, a second ablation study was conducted to assess the addition of a quantitative T2 map in model training. The data were evaluated using an 8-fold leave-5-out cross-validation method. Pixel-wise biventricular O₂Sat values predicted from test data were averaged and compared against reference values. DeepOx RV results were also compared to model-based MR oximetry [3,4].

Results: Table 1 shows R² for the ablation studies; training DeepOx with T2-prepared images alone and Huber loss yielded the best performance, with no added benefit from T2 maps. DeepOx achieved an average R² of 0.95 and MAE of 0.022 on test sets for RV and LV O₂Sat predictions. With respect to RV O₂Sat, DeepOx performed better than prior model-based oximetry (R²: 0.60 vs. 0.40; MAE: 0.031 vs. 0.041; Figure 2a-b). Bland-Altman plots (Figure 2c-d) demonstrate slightly reduced bias and tighter limits of agreement for DeepOx. Within an accuracy of 5% and 10% of the ground truth, DeepOx reached 82% and 97% accuracy, respectively, compared to 75% and 94% for model-based oximetry.

Conclusion: DeepOx potentially allows accurate and model-free, non-invasive MR estimation of intracardiac O₂Sat. Non-reliance on blood sampling for hematocrit and a pulse-oximeter-based LV reference O₂ saturation for patient-specific inputs further improves upon previously shown model-based MR oximetry. With additional training and validation in larger multi-center and more diverse datasets with different imaging planes and patient conditions, DeepOx could further support integration of intracardiac O₂Sat into CMR assessment.

Table 1: Ablation Study

Results: Study 1 evaluated loss functions in DeepOx training. Study 2 evaluated the inclusion of a T2 map in training. R2 results are shown for combined RV and LV test data, RV and LV test data independently, and for 5% accuracy compared to the reference.

	Ablation study #1 (Assess loss function)			Ablation study #2 (Assess input channels
Constant	T2-prepared images			Huber Loss
Variable	MSE	MAE	Huber (δ=0.05)	T2-prepared images	T2 map	T2-prepared images + T2 map
RV & LV Test	0.89	0.92	0.95	0.95	0.92	0.94
RV Test	0.22	0.34	0.60	0.60	0.44	0.55
LV Test	-3.47	-0.21	0.21	0.21	-1.50	-0.45
5% Accuracy	0.86	0.84	0.91	0.91	0.82	0.86

Figure 1: U-NET architecture used to train the model. Six T2-prepared single-shot images are provided as input channels. The output shows the predicted DeepOx results in comparison to the ground truth image generated by the right heart catheter and pulse oximeter-based reference O2Sat values for the right and left ventricles, respectively.
Figure 2: Performance evaluation of DeepOx O2Sat predictions compared to reference O₂ saturation measurements. (a) Scatter plot of chamber-averaged DeepOx predictions versus reference O₂ saturation (catheter for right ventricle and pulse oximeter for left ventricle) for training (blue) and testing (red) datasets.(b) Comparison of RV O2 saturation results between DeepOx (red) and model-based oximetry (green) with respect to reference catheter-based measurements. Bland–Altman plots for (c) DeepOx and (d) model-based oximetry.