Background: For both hospitals and patients it would be beneficial if the scan time of MR images could be reduced. At the moment, Compressed Sensing (CS) is introduced to reduce the scan time, however, new methods are developed such as a deep learning method, called the Recurrent Inference Machine (RIM). In this study the effect of reconstructing undersampled MRI images with lesions, using the RIM and CS, was evaluated. In data of a healthy control, lesions were simulated. Evaluation is done by checking if the lesion has the correct intensity and shape after reconstruction of undersampled data. Methods: In raw data of a healthy control lesions were simulated. To test the RIM and CS, the images with lesions where first undersampled 4x, 6x, 8x and 10x. After undersampling, the images were reconstructed with both RIM and CS. First, the peak intensity difference was measured between the reference image (with simulated lesions) and reconstructed images for both RIM and CS. Second, one lesion was undersampled ten times with different undersampling masks creating different noise, for 3 different acceleration factors (4x, 6x, 8x). These lesions were reconstructed with both RIM and CS. The maximum intensity difference between reference and reconstructed image was measured and averaged over the ten different undersampled images. Results: In total seven different lesions were simulated in a healthy control with different intensities varying between 10\% and 100\% of the GM-lesion intensity in a FLAIR weighted scan. The intensities of all lesions were more accurately reconstructed with the RIM compared to CS at higher acceleration factors: the average intensity per lesions after 10 times reconstruction with RIM was more equal to the correct intensity compared to the reconstruction with CS. Conclusion: The RIM shows robust and accurate results on data with simulated lesions. Moreover, the RIM outperformed CS on data that was more undersampled. Therefore, the RIM may be used for reconstruction of MRI data that is acquired with shorter acquisition time. And since the reconstruction time is better, it could replace CS in the future. However, before the RIM could be used, further evaluations on actual patient data are needed.