Summary: After the successful clinical introduction of PET/CT, a novel hybrid imaging technology combining PET with the versatile attributes of MRI is emerging. At the Forschungszentrum Jülich, one of four prototypes available worldwide combining a commercial 3T MRI with a newly developed BrainPET insert has been installed, allowing simultaneous data acquisition with PET and MRI. The BrainPET is equipped with LSO crystals of 2.5 mm width and Avalanche photodiodes (APD) as readout electronics. Here we report on some performance characteristics obtained by phantom studies and also on the initial BrainPET studies on various patients as compared with a conventional HR+ PET-only scanner. Material, methods: The radiotracers [18F]-fluoroethyl- tyrosine (FET), [11C]-flumazenil and [18F]-FP-CIT were applied. Results: Comparing komthe PET data obtained with the BrainPET to those of the HR+ scanner demonstrated the high image quality and the superior resolution capability of the BrainPET. Furthermore, it is shown that various MR images of excellent quality could be acquired simultaneously with BrainPET scans without any relevant artefacts. Discussion, conclusion: Initial experiences with the hybrid MRI/BrainPET indicate a promising basis for further developments of this unique technique allowing simultaneous PET imaging combined with both anatomical and functional MRI.

BACKGROUND AND PURPOSE: Malignant melanoma is an aggressive skin cancer in which brain metastases are common. Our aim was to establish and evaluate a deep learning model for fully automated detection and segmentation of brain metastases in patients with malignant melanoma using clinical routine MR imaging. MATERIALS AND METHODS: Sixty-nine patients with melanoma with a total of 135 brain metastases at initial diagnosis and available multiparametric MR imaging datasets (T1-/T2-weighted, T1-weighted gadolinium contrast-enhanced, FLAIR) were included. A previously established deep learning model architecture (3D convolutional neural network; DeepMedic) simultaneously operating on the aforementioned MR images was trained on a cohort of 55 patients with 103 metastases using 5-fold cross-validation. The efficacy of the deep learning model was evaluated using an independent test set consisting of 14 patients with 32 metastases. Manual segmentations of metastases in a voxelwise manner (T1-weighted gadolinium contrast-enhanced imaging) performed by 2 radiologists in consensus served as the ground truth. RESULTS: After training, the deep learning model detected 28 of 32 brain metastases (mean volume, 1.0 [SD, 2.4] cm³) in the test cohort correctly (sensitivity of 88%), while false-positive findings of 0.71 per scan were observed. Compared with the ground truth, automated segmentations achieved a median Dice similarity coefficient of 0.75. CONCLUSIONS: Deep learning–based automated detection and segmentation of brain metastases in malignant melanoma yields high detection and segmentation accuracy with false-positive findings of <1 per scan.

Simultaneous, hybrid MR-PET is expected to improve PET image resolution in the plane perpendicular to the static magnetic field of the scanner. Previous papers have reported this either by simulation or experiment with simple sources and detector arrangements. Here, we extend those studies using a realistic brain phantom in a recently installed MR-PET system comprising a 9.4 T MRI-scanner and an APD-based BrainPET insert in the magnet bore. Point and line sources and a 3D brain phantom were filled with ¹⁸F (low-energy positron emitter), ⁶⁸Ga (medium energy positron emitter) or ¹²⁰I, a non-standard positron emitter (high positron energies of up to 4.6 MeV). Using the BrainPET insert, emission scans of the phantoms were recorded at different positions inside and outside the magnet bore such that the magnetic field was 0 T, 3 T, 7 T or 9.4 T. Brain phantom images, with the ‘grey matter’ compartment filled with ¹⁸F, showed no obvious resolution improvement with increasing field. This is confirmed by practically unchanged transaxial FWHM and ‘grey/white matter’ ratio values between at 0T and 9.4T. Field-dependent improvements in the resolution and contrast of transaxial PET images were clearly evident when the brain phantom was filled with ⁶⁸Ga or ¹²⁰I. The grey/white matter ratio increased by 7.3% and 16.3%, respectively. The greater reduction of the FWTM compared to FWHM in ⁶⁸Ga or ¹²⁰I line-spread images was in agreement with the improved contrast of ⁶⁸Ga or ¹²⁰I images. Notwithstanding elongations seen in the z-direction of ⁶⁸Ga or ¹²⁰I point source images acquired in foam, brain phantom images show no comparable extension. Our experimental study confirms that integrated MR-PET delivers improved PET image resolution and contrast for medium- and high-energy positron emitters even though the positron range is reduced only in directions perpendicular to the magnetic field.

Kirsten Rat Sarcoma viral oncogene homolog (KRAS) is a frequently occurring mutation in non-small-cell lung cancer (NSCLC) and influences cancer treatment and disease progression. In this study, a machine learning (ML) pipeline was applied to radiomic features extracted from public and internal CT images to identify KRAS mutations in NSCLC patients. Both datasets were analyzed using parametric (t test) and non-parametric statistical tests (Mann–Whitney U test) and dimensionality reduction techniques. Afterwards, the proposed ML pipeline was applied to both datasets using a five-fold cross-validation on the training set (70/30 train/test split) before being validated on the other dataset. The results show that the radiomic features are significantly different (Mann–Whitney U test; p < 0.05) between the two datasets, despite the use of identical feature extraction methods. Model transferability is therefore difficult to achieve, which became evident during external testing (F1 score = 0.41). Oversampling, undersampling, clustering and harmonization techniques were applied to balance and harmonize the datasets, but did not improve the classification of KRAS mutation presence. In general, due to only a single moderate result (highest test F1 score = 0.67), the accuracy of KRAS prediction is not sufficient for clinical application. In future work, the complexity of KRAS mutation might be addressed by taking submutations into consideration. Larger multicentric datasets with balanced tumor stages, including multi-scanner datasets, seem to be necessary for building robust predictive models.

Currently, the medical field is witnessing an increase in the use of machine learning techniques. Supervised learning methods adopted in classification, prediction, and segmentation tasks for medical images always experience decreased performance when the training and testing datasets do not follow the independent and identically distributed assumption. These distribution shift situations seriously influence machine learning applications’ robustness, fairness, and trustworthiness in the medical domain. Hence, in this article, we adopt the CycleGAN (generative adversarial network) method to cycle train the computed tomography data from different scanners/manufacturers. It aims to eliminate the distribution shift from diverse data terminals based on our previous work [ 14 ]. However, due to the model collapse problem and generative mechanisms of the GAN-based model, the images we generated contained serious artifacts. To remove the boundary marks and artifacts, we adopt score-based diffusion generative models to refine the images voxel-wisely. This innovative combination of two generative models enhances the quality of data providers while maintaining significant features. Meanwhile, we use five paired patients’ medical images to deal with the evaluation experiments with structural similarity index measure metrics and the segmentation model’s performance comparison. We conclude that CycleGAN can be utilized as an efficient data augmentation technique rather than a distribution-shift-eliminating method. In contrast, the denoising diffusion the denoising diffusion model is more suitable for dealing with the distribution shift problem aroused by the different terminal modules. The limitation of generative methods applied in medical images is the difficulty in obtaining large and diverse datasets that accurately capture the complexity of biological structure and variability. In our following research, we plan to assess the initial and generated datasets to explore more possibilities to overcome the above limitation. We will also incorporate the generative methods into the federated learning architecture, which can maintain their advantages and resolve the distribution shift issue on a larger scale.