1344575 - Fully Automatic Scar Segmentation on LGE Images with Reduced Contrast Agent Dose

Assessment of the extent of late gadolinium enhancement (LGE) in patients with myocardial infarction (MI) is a key biomarker for treatment and prognosis. Concerns about the safety of gadolinium-based contrast agents (GBCA) [1] have led to an interest in reducing the amount of administered GBCA. We propose a deep learning-based approach for scar segmentation and quantification on LGE images with reduced GBCA doses.

Methods:

Forty-one patients with MI underwent a two-stage LGE imaging protocol on a 1.5T scanner (MAGNETOM AvantoFit, Siemens Healthcare, Erlangen, Germany). A low-dose PSIR LGE image in short-axis orientation was acquired after injection of 0.1 mmol/kg Gadoteridol (imaged ~3-5 minutes post-injection). A second, full-dose PSIR image was obtained after injection of additional 0.05 mmol/kg GBCA (imaged ~15-20 minutes after the first injection). An additional 27 patients underwent a CMR exam with only the full-dose PSIR. Myocardial borders were contoured by a reader with three years of experience and the scar area was determined using the full width at half maximum (FWHM) approach. Image registration of low- and full-dose images was performed based on the myocardial borders. In addition, the publicly available EMIDEC [2] dataset was used for training.
The training procedure is illustrated in Fig. 1. First, a U-Net was trained to segment the left ventricle, allowing to crop the images around the LV for subsequent processing. A second U-Net was then trained to segment healthy and scarred myocardium in the cropped images. Only images acquired with the regular GBCA dose were used so far due to higher data availability. The second network was then further trained and finetuned using the low-dose images acquired with the reduced GBCA dose. The ground-truth segmentation masks used in this training stage were taken from the corresponding full-dose images.

Results:

The network was evaluated on low-dose images from eight patients that underwent the two-stage imaging protocol. The corresponding segmentation masks at full dose were used as reference. Mean Dice similarity coefficients (DSC) of 80.0% and 36.7% were achieved for healthy myocardium and scar, respectively. The mean error of the relative scar area (difference of scar areas divided by total myocardium areas) was -2.82%. Exemplary segmentation results are given in Fig. 2.

For comparison, the FWHM method was applied on the low-dose images and achieved a DSC of 59.0% for scar and a mean scar area error of +3.14%. Bland-Altman plots are given in Fig. 3.

Conclusion:

The proposed method allows to accurately quantify LGE on images with reduced contrast agent dose without the need for manual contouring. Although the DSC indicates a lower segmentation quality, this metric may be significantly influenced by slightly inaccurate segmentations of small scars. In terms of quantification, the network consistently underestimated the scar size, but still achieved better results than directly applying the FWHM method on the low-dose images.