Complexities of deep learning-based undersampled MR image reconstruction

Artificial intelligence has opened a new path of innovation in magnetic resonance (MR) image reconstruction of undersampled k-space acquisitions. This review offers readers an analysis of the current deep learning-based MR image reconstruction methods. The literature in this field shows exponential growth, both in volume and complexity, as the capabilities of machine learning in solving inverse problems such as image reconstruction are explored. We review the latest developments, aiming to assist researchers and radiologists who are developing new methods or seeking to provide valuable feedback. We shed light on key concepts by exploring the technical intricacies of MR image reconstruction, highlighting the importance of raw datasets and the difficulty of evaluating diagnostic value using standard metrics.Relevance statement Increasingly complex algorithms output reconstructed images that are difficult to assess for robustness and diagnostic quality, necessitating high-quality datasets and collaboration with radiologists.Key points• Deep learning-based image reconstruction algorithms are increasing both in complexity and performance.• The evaluation of reconstructed images may mistake perceived image quality for diagnostic value.• Collaboration with radiologists is crucial for advancing deep learning technology.

Balanced Steady-State Free Precession Techniques Improve Detection of Residual Germ Cell Tumor for Treatment Planning

BACKGROUND AND PURPOSE

Identification of a partial/complete chemotherapy response in pediatric patients with intracranial germ cell tumors is clinically important for radiation treatment and management. Partial/complete response is conventionally determined on postcontrast MR imaging sequences. The purpose of this study was to assess the diagnostic utility of a balanced steady-state free precession sequence as an adjunct to standard MR imaging sequences for the detection of residual tumor in pediatric patients on postchemoreduction pre-radiation planning MR imaging.

MATERIALS AND METHODS

This was a retrospective study of pediatric patients with intracranial germ cell tumors undergoing postchemotherapy, preradiotherapy MR imaging. Patients underwent 1.5T or 3T MR imaging with pre- and postcontrast T1WIs, T2WIs, and a balanced steady-state free precession sequence. Two neuroradiologists independently reviewed standard MR imaging sequences without the balanced steady-state free precession sequence, then with the balanced steady-state free precession sequence 1 week later. Assessment for partial/complete response was determined using Response Assessment in Neuro-Oncology criteria. A 5-point Likert scale scored the diagnostic confidence of the neuroradiologist rating each study without/with the balanced steady-state free precession sequence. Rates of residual disease concordance and diagnostic confidence levels without/with the balanced steady-state free precession sequence were calculated.

RESULTS

Thirty-nine patients were included with 31 males and 8 females (mean age, 14.15 ± 4.26 years). Thirty-one patients had single-site disease; 8 patients had multisynchronous disease (47 sites in total). Compared to review of the standard MR sequences alone, the addition of the balanced steady state free precession sequence resulted in higher rates of tumor partial response categorization and greater diagnostic confidence levels (P < .001, P < .001).

CONCLUSIONS

The balanced steady-state free precession sequence improves detection of residual chemotherapy-reduced intracranial germ cell tumors and increases diagnostic confidence of the neuroradiologist. The balanced steady-state free precession sequence may be an important adjunct to the standard MR imaging protocol for radiation planning.

Validating subject-specific RF and thermal simulations in the calf muscle using MR-based temperature measurements

PURPOSE

Ongoing discussions occur to translate the safety restrictions on MR scanners from specific absorption rate (SAR) to thermal dose. Therefore, this research focuses on the accuracy of thermal simulations in human subjects during an MR exam, which is fundamental information in that debate.

METHODS

Radiofrequency (RF) heating experiments were performed on the calves of 13 healthy subjects using a dedicated transmit-receive coil while monitoring the temperature with proton resonance frequency shift (PRFS) thermometry. Subject-specific models and one generic model were used for electromagnetic and thermal simulations using Pennes' bioheat equation, with the blood equilibration constant equaling zero. The simulations were subsequently compared with the experimental results.

RESULTS

The mean B1+ equaled 15 µT in the center slice of all volunteers, and 95% of the voxels had errors smaller than 2.8 µT between the simulation and measurement. The intersubject variation in RF power to achieve the required B1+ was 11%. The resulting intersubject variation in median temperature rise was 14%. Thermal simulations underestimated the median temperature increase on average, with 34% in subject-specific models and 28% in the generic model.

CONCLUSIONS

Although thermal measures are directly coupled to tissue damage and therefore suitable for RF safety assessment, insecurities in the applied thermal modeling limit their estimation accuracy. Magn Reson Med 77:1691-1700, 2017. © 2016 International Society for Magnetic Resonance in Medicine.