Mahdjoub Hamdi

PET/MRI is a highly versatile and effective imaging modality in pediatric populations that requires dedicated study preparation, optimized imaging protocols, and structured interpretation to reach its full potential.

Purpose: Challenges in PET/MRI quantitative accuracy for neurological uses arise from PET attenuation correction accuracy. We proposed and evaluated an automatic pipeline to assess the quantitative accuracy of four MRI-derived PET AC methods using analytically simulated PET brain lesions and ROIs as ground truth for PET activity. Methods: Our proposed pipeline, integrating a synthetic lesion insertion tool and the FreeSurfer neuroimaging framework, inserts simulated spherical and brain ROIs into PET projection space, reconstructing them via four PET MRAC techniques. Utilizing an 11-patient brain PET dataset, we compared the quantitative accuracy of four MRACs (DIXON, DIXONbone, UTE AC, and DL-DIXON) against the gold standard PET CTAC, evaluating MRAC to CTAC activity bias in spherical lesions and brain ROIs with and without background activity against original (lesion free) PET reconstructed images. Results: The proposed pipeline yielded accurate results for spherical lesions and brain ROIs, adhering to the MRAC to CTAC pattern of original brain PET images. Among the MRAC methods, DIXON AC exhibited the highest bias, followed by UTE, DIXONBone, and DL-DIXON showing the least. DIXON, DIXONbone, UTE, and DL-DIXON showed MRAC to CTAC biases of − 5.41%, − 1.85%, − 2.74%, and 0.08% respectively for ROIs inserted in background activity; − 7.02%, − 2.46%, − 3.56%, and − 0.05% for lesion ROIs without background; and − 6.82%, − 2.08%, − 2.29%, and 0.22% for the original brain PET images’ 16 FreeSurfer brain ROIs. Conclusion: The proposed pipeline delivers accurate results for synthetic spherical lesions and brain ROIs, with and without background activity consideration, enabling the evaluation of new attenuation correction approaches without utilizing measured PET emission data. Additionally, it offers a consistent method to generate realistic lesion ROIs, potentially applicable in assessing further PET correction techniques.

Purpose PET/MRI quantitative accuracy for neurological applications is challenging due to accuracy of the PET attenuation correction. In this work, we proposed and evaluated an automatic pipeline for assessing the quantitative accuracy of four different MRI = based attenuation correction (PET MRAC) approaches. Methods The proposed pipeline consists of a synthetic lesion insertion tool and the FreeSurfer neuroimaging analysis framework. The synthetic lesion insertion tool is used to insert simulated spherical, and brain regions of interest (ROI) into the PET projection space and reconstructed with four different PET MRAC techniques, while FreeSurfer is used to generate brain ROIs from T1 weighted MRI image. Using a cohort of 11 patients' brain PET dataset, the quantitative accuracy of four MRAC(s), which are: DIXON AC, DIXONbone AC, UTE AC, and Deep learning trained with DIXON AC, named DL-DIXON AC, were compared to the PET-based CT attenuation correction (PET CTAC). MRAC to CTAC activity bias in spherical lesions and brain ROIs were reconstructed with and without background activity and compared to the original PET images. Results The proposed pipeline provides accurate and consistent results for inserted spherical lesions and brain ROIs inserted with and without considering the background activity and following the same MRAC to CTAC pattern as the original brain PET images. As expected, the DIXON AC showed the highest bias; the second was for the UTE, then the DIXONBone, and the DL-DIXON with the lowest bias. For simulated ROIs inserted in the background activity, DIXON showed a -4.65% MRAC to CTAC bias, 0.06% for the DIXONbone, -1.70% for the UTE, and − 0.23% for the DL-DIXON. For lesion ROIs inserted without background activity, DIXON showed a -5.21%, -1% for the DIXONbone, -2.55% for the UTE, and − 0.52 for the DL-DIXON. For MRAC to CTAC bias calculated using the same 16 FreeSurfer brain ROIs in the original brain PET reconstructed images, a 6.87% was observed for the DIXON, -1.83% for DIXON bone, -3.01% for the UTE, and − 0.17% for the DL-DIXON. Conclusion The proposed pipeline provides accurate and consistent results for synthetic spherical lesions and brain ROIs inserted with and without considering the background activity; hence a new attenuation correction approach can be evaluated without using measured PET emission data.

Historically, patient datasets have been used to develop and validate various reconstruction algorithms for PET/MRI and PET/CT. To enable such algorithm development, without the need for acquiring hundreds of patient exams, in this paper we demonstrate a deep learning technique to generate synthetic but realistic whole-body PET sinograms from abundantly-available whole-body MRI. Specifically, we use a dataset of 56 18F-FDG-PET/MRI exams to train a 3D residual UNet to predict physiologic PET uptake from whole-body T1-weighted MRI. In training we implemented a balanced loss function to generate realistic uptake across a large dynamic range and computed losses along tomographic lines of response to mimic the PET acquisition. The predicted PET images are forward projected to produce synthetic PET time-of-flight (ToF) sinograms that can be used with vendor-provided PET reconstruction algorithms, including using CT-based attenuation correction (CTAC) and MR-based attenuation correction (MRAC). The resulting synthetic data recapitulates physiologic 18F-FDG uptake, e.g. high uptake localized to the brain and bladder, as well as uptake in liver, kidneys, heart and muscle. To simulate abnormalities with high uptake, we also insert synthetic lesions. We demonstrate that this synthetic PET data can be used interchangeably with real PET data for the PET quantification task of comparing CT and MR-based attenuation correction methods, achieving ≤7.6% error in mean-SUV compared to using real data. These results together show that the proposed synthetic PET data pipeline can be reasonably used for development, evaluation, and validation of PET/MRI reconstruction methods.

Histopathological analysis of whole-slide images is the gold standard technique for diagnosis of lung cancer and classifying it into types and subtypes by specialized pathologists. This labor-based approach is time and effort consuming, which led to development of automatic approaches to assist in reducing the time and effort. Deep learning is a supervised classification approach that is well adapted for automatic classification of histopathological images. We aimed to develop a deep learning-based approach for lung adenocarcinoma pattern classification and generalize the proposed approach to the classification of the major non-small cell lung cancer types. Three publicly available datasets were used in this study. A deep learning approach for histopathological image analysis using convolutional neural networks was developed and incorporated into automatic pipelines to accurately classify the predominant patterns on the whole-slide images level and non-small cell lung cancer types on patch-level. The models were evaluated using the confusion matrix to perform an error analysis and the classification report to compute F1-score, recall and precision. As results, the three models have shown an excellent performance with best combination of hyper-parameters for training models. First and second models predicted adenocarcinoma predominant patterns on two different datasets with an accuracy, respectively, of 96.15% and 89.51%. The third model has exceeded an accuracy of 99.72% in classifying major non-small cell lung cancer types. The proposed deep learning-based lung cancer classification approach can be used to assist pathologists in identifying of lung adenocarcinomas patterns.