About

Sergios Gatidis is an Associate Professor of Radiology at Stanford University and is affiliated with the Center for Artificial Intelligence in Medicine & Imaging (AIMI). His work focuses on the application of artificial intelligence in the field of medicine and imaging, contributing to the advancement of AI-driven healthcare solutions. As part of AIMI, he is involved in research initiatives that aim to improve medical imaging analysis and develop innovative AI methodologies for healthcare. His role includes engaging in collaborative projects, educational activities, and leadership within the center dedicated to integrating AI technologies into medical practice.

Research topics

• Computer Science
• Medicine
• Nuclear medicine
• Information Retrieval
• Radiology
• Natural Language Processing
• Artificial Intelligence
• Medical physics

Selected publications

• PSMA-FDG-PET-CT-Lesions

  FDAT · 2026-04-26

  otherOpen access1st authorCorresponding

  Introduction A publicly available dataset of annotated Positron Emission Tomography/Computed Tomography (PET/CT) studies. 1014 whole body Fluorodeoxyglucose (FDG)-PET/CT studies (900 patients) and 597 prostate-specific membrane antigen (PSMA)-PET/CT studies (378 patients) acquired between 2014 and 2022 were included. The FDG cohort comprises 501 patients diagnosed with histologically proven malignant melanoma, lymphoma, or lung cancer, along with 513 negative control patients. The PSMA cohort includes pre- and/or post-therapeutic PET/CT images of male individuals with prostate carcinoma, encompassing images with (537) and without PSMA-avid tumor lesions (60). Notably, the training datasets exhibit distinct age distributions: the FDG UKT cohort spans 570 male patients (mean age: 60; std: 16) and 444 female patients (mean age: 58; std: 16), whereas the PSMA LMU cohort tends to be older, with 378 male patients (mean age: 71; std: 8). Additionally, there are variations in imaging conditions between the FDG Tübingen and PSMA Munich cohorts, particularly regarding the types and number of PET/CT scanners utilized for acquisition. The PSMA Munich dataset was acquired using three different scanner types (Siemens Biograph 64-4R TruePoint, Siemens Biograph mCT Flow 20, and GE Discovery 690), whereas the FDG Tübingen dataset was acquired using a single scanner (Siemens Biograph mCT). Structure and usage The data is organized in the nnUNet structure: |--- imagesTr |--- tracer_patient1_study1_0000.nii.gz (CT image resampled to PET) |--- tracer_patient1_study1_0001.nii.gz (PET image in SUV) |--- ... |--- labelsTr |--- tracer_patient1_study1.nii.gz (manual annotations of tumor lesions) |--- dataset.json (nnUNet dataset description) |--- dataset_fingerprint.json (nnUNet dataset fingerprint) |--- splits_final.json (reference 5fold split) |--- psma_metadata.csv (metadata csv for psma) |--- fdg_metadata.csv (original metadata csv for fdg) We demonstrate how this dataset can be used for deep learning-based automated analysis of PET/CT data and provide the trained deep learning model: www.autopet.org PET/CT acquisition protocol FDG dataset: Patients fasted at least 6 h prior to the injection of approximately 350 MBq 18F-FDG. Whole-body PET/CT images were acquired using a Biograph mCT PET/CT scanner (Siemens, Healthcare GmbH, Erlangen, Germany) and were initiated approximately 60 min after intravenous tracer administration. Diagnostic CT scans of the neck, thorax, abdomen, and pelvis (200 reference mAs; 120 kV) were acquired 90 sec after intravenous injection of a contrast agent (90-120 ml Ultravist 370, Bayer AG) or without contrast agent (in case of existing contraindications). PET Images were reconstructed iteratively (three iterations, 21 subsets) with Gaussian post-reconstruction smoothing (2 mm full width at half-maximum). Slice thickness on contrast-enhanced CT was 2 or 3 mm. PSMA dataset: Examinations were acquired on different PET/CT scanners (Siemens Biograph 64-4R TruePoint, Siemens Biograph mCT Flow 20, and GE Discovery 690). The imaging protocol mainly consisted of a diagnostic CT scan from the skull base to the mid-thigh using the following scan parameters: reference tube current exposure time product of 143 mAs (mean); tube voltage of 100kV or 120 kV for most cases, slice thickness of 3 mm for Biograph 64 and Biograph mCT, and 2.5 mm for GE Discovery 690 (except for 3 cases with 5 mm). Intravenous contrast enhancement was used in most studies (571), except for patients with contraindications (26). The whole-body PSMA-PET scan was acquired on average around 74 minutes after intravenous injection of 246 MBq 18F-PSMA (mean, 369 studies) or 214 MBq 68Ga-PSMA (mean, 228 studies), respectively. The PET data was reconstructed with attenuation correction derived from corresponding CT data. For GE Discovery 690 the reconstruction process employed a VPFX algorithm with voxel size 2.73 mm × 2.73 mm × 3.27 mm, for Siemens Biograph mCT Flow 20 a PSF+TOF algorithm (2 iterations, 21 subsets) with voxel size 4.07 mm × 4.07 mm × 3.00 mm, and for Siemens Biograph 64-4R TruePoint a PSF algorithm (3 iterations, 21 subsets) with voxel size 4.07 mm × 4.07 mm × 5.00 mm. Annotation FDG PET/CT training and test data from UKT was annotated by a Radiologist with 10 years of experience in Hybrid Imaging and experience in machine learning research. FDG PET/CT test data from LMU was annotated by a radiologist with 8 years of experience in hybrid imaging. PSMA PET/CT training and test data from LMU as well as PSMA PET/CT test data from UKT was annotated by a single reader and reviewed by a radiologist with 5 years of experience in hybrid imaging. The following annotation protocol was defined:Step 1: Identification of tracer-avid tumor lesions by visual assessment of PET and CT information together with the clinical examination reports.Step 2: Manual free-hand segmentation of identified lesions in axial slices. Versions v1: Initial submissionv2: Corrected segmentation masksv3: QIBA-aligned SUV calculations of the PET imaging following the codes here

  PublisherDOI

• FDG-PET-CT-Lesions

  FDAT · 2026-04-27

  otherOpen access1st authorCorresponding

  Introduction A publicly available dataset of annotated Positron Emission Tomography/Computed Tomography (PET/CT) studies. 1014 whole body Fluorodeoxyglucose (FDG)-PET/CT datasets (501 studies of patients with malignant lymphoma, melanoma and non small cell lung cancer (NSCLC) and 513 studies without PET-positive malignant lesions (negative controls)) acquired between 2014 and 2018 were included. All examinations were acquired on a single, state-of-the-art PET/CT scanner. The imaging protocol consisted of a whole-body FDG-PET acquisition and a corresponding diagnostic CT scan. All FDG-avid lesions identified as malignant based on the clinical PET/CT report were manually segmented on PET images in a slice-per-slice (3D) manner. We provide the anonymized NIfTI files of all studies as well as the corresponding NIfTI segmentation masks. In addition, we provide scripts for image processing and conversion to different file formats (DICOM, mha, hdf5). Primary diagnosis, age and sex are provided as non-imaging information. Structure and usage The data is organized in the following structure: |--- Patient 1 |--- Study 1 |--- SUV.nii.gz (PET image in SUV) |--- CTres.nii.gz (CT image resampled to PET) |--- CT.nii.gz (Original CT image) |--- SEG.nii.gz (Manual annotations of tumor lesions) |--- PET.nii.gz (Original PET image as activity counts) |--- Study 2 (Potential 2nd visit of same patient) |--- ... |--- Patient 2 |--- ... |--- fdg_metadata.csv (metadata csv for studies) We demonstrate how this dataset can be used for deep learning-based automated analysis of PET/CT data and provide the trained deep learning model: www.autopet.org PET/CT acquisition protocol Patients fasted at least 6 h prior to the injection of approximately 350 MBq 18F-FDG. Whole-body PET/CT images were acquired using a Biograph mCT PET/CT scanner (Siemens, Healthcare GmbH, Erlangen, Germany) and were initiated approximately 60 min after intravenous tracer administration. Diagnostic CT scans of the neck, thorax, abdomen and pelvis (200 reference mAs; 120 kV) were acquired 90 sec after intravenous injection of a contrast agent (90–120 ml Ultravist 370, Bayer AG). PET Images were reconstructed iteratively (three iterations, 21 subsets) with Gaussian post-reconstruction smoothing (2 mm full width at half-maximum). Slice thickness on contrast-enhanced CT was 2 or 3 mm. Annotation Two experts annotated training and test data: At the University Hospital Tübingen, a Radiologist with 10 years of experience in Hybrid Imaging and experience in machine learning research annotated all data. At the University Hospital of the LMU in Munich, a Radiologist with 5 years of of experience in Hybrid Imaging and experience in machine learning research annotated all data. The following annotation protocol was defined:Step 1: Identification of FDG-avid tumor lesions by visual assessment of PET and CT information together with the clinical examination reports.Step 2: Manual free-hand segmentation of identified lesions in axial slices. Versions v1: Initial submissionv2: QIBA-aligned SUV calculations of the PET imaging following the codes here

  PublisherDOI

• PSMA-PET-CT-Lesions

  FDAT · 2026-04-26

  datasetOpen access

  Introduction We provide a large, annotated dataset of 597 whole-body PSMA-PET/CT studies from 378 male patients with suspected or diagnosed prostate carcinoma to support developing and benchmarking machine learning (ML) models for automated quantitative PET/CT analysis. Alongside the FDG-PET/CT dataset, this dataset addresses the scarcity of publicly available, high-quality annotated PET/CT data. The FDG and PSMA-PET/CT datasets were jointly provided as training data for developing ML models in the autoPET III and IV Grand Challenges for automated lesion segmentation in whole-body PET/CT. Data Acquisition Scans were conducted at LMU University Hospital, LMU Munich, between 2014 and 2022 using three clinical PET/CT scanners: Siemens Biograph mCT Flow 20, Siemens Biograph 64-4R TruePoint, and GE Discovery 690. 539 studies contain at least one PSMA-avid tumor lesion, 58 studies do not contain any PSMA-avid tumor lesion. The imaging protocol consisted of a diagnostic CT scan usually from the skull base to the mid-thigh with the following scan parameters: reference tube current exposure time product of 143 mAs (mean); tube voltage of 120 kV or 100 kV for most cases (range: [80, 140] kV), slice thickness of 2.5 - 5.0 mm (mean: 2.82 mm), and x-y resolution of mainly 0.98 mm. Intravenous contrast enhancement was used in most studies, except for patients with contraindications (26 studies). The whole-body PSMA-PET scan was acquired on average 74 minutes after intravenous injection of 246 MBq 18F-PSMA (mean, 369 studies) or 214 MBq 68Ga-PSMA (mean, 228 studies), respectively. The PET data was reconstructed with attenuation correction derived from corresponding CT data using standard, vendor-provided image reconstruction algorithms with a slice thickness ranging from 3.0 - 5.0 mm (mean: 3.49 mm) and x-y resolution ranging from 2.73 - 4.07 mm (mean: 3.56 mm). Data Annotation All PSMA-avid tumor lesions, including the primary tumor and/or all metastases, were manually segmented on the PET images by a single reader with 3 years of experience in hybrid imaging using dedicated software (mint Medical, Heidelberg, Germany) and validated by board-certified medical imaging experts with 4 years and >10 years of experience in hybrid imaging. Tumor lesions with significantly increased PSMA expression were segmented in 3D space by drawing circular VOIs, in which voxels with uptake values above a user-defined threshold were pre-segmented automatically and then manually corrected slice by slice, resulting in 3D binary segmentation masks. Data Processing For DICOM-to-NIfTI conversion, CT volumes were resampled to match the size and resolution of the corresponding PET volume, PET voxel values were normalized to standardized uptake values (SUV) based on body mass. In addition, patient metadata was extracted from imaging DICOM tags and saved in a CSV file: patient age at imaging in years, PET/CT manufacturer and model name, PET radionuclide, and use of CT contrast agent. Information on radionuclides and the use of CT contrast agents was visually reviewed and validated by a radiologist with 10 years of experience in hybrid imaging. Each study is uniquely identified by an anonymized case identifier number and the study date. The study date was shifted by a global patient-level offset, such that differences between the study dates of a patient are conserved. Data structure The NIfTI dataset is organized in nnU-Net structure. |--- imagesTr |--- <tracer>_<patient_1>_<study_1>_0000.nii.gz (CT image resampled to PET) |--- <tracer>_<patient_1>_<study_1>_0001.nii.gz (PET image in SUV) |--- ... |--- labelsTr |--- <tracer>_<patient_1>_<study_1>.nii.gz (SEG mask) |--- ... |--- dataset.json (nnUNet dataset description) |--- dataset_fingerprint.json (nnUNet dataset fingerprint) |--- splits_final.json (reference 5-fold split) |--- psma_metadata.csv (metadata csv for psma) Usage The dataset can be used for training deep learning models for automated lesion segmentation in whole-body PET/CT: www.autopet.org Versions v1: Initial release v2: Corrected segmentation masks in 5 studies from 3 patients. Lesion segmentations were updated in 3 studies, and added to 2 studies that previously had none. Updated files: labelsTr/psma_4da96443cf212c5f_2020-08-31.nii.gz labelsTr/psma_4da96443cf212c5f_2022-01-15.nii.gz labelsTr/psma_4da96443cf212c5f_2022-05-14.nii.gz labelsTr/psma_5eb9920ce854b7a2_2019-03-29.nii.gz labelsTr/psma_faa4c90c3d2d53a3_2019-08-10.nii.gz v3: QIBA-aligned SUV calculations of the PET imaging following the codes here

  PublisherDOI

• Pediatric PET/MRI: Imaging Techniques, Indications, and Clinical Implementation

  Radiographics · 2025-10-16 · 2 citations

  article1st authorCorresponding

  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.

  PublisherDOI

• Exploration of Whole-Body Anatomy in the German National Cohort (NAKO): 3D Segmentation of 55 Structures in 28,969 MRI Scans

  Proceedings on CD-ROM - International Society for Magnetic Resonance in Medicine. Scientific Meeting and Exhibition/Proceedings of the International Society for Magnetic Resonance in Medicine, Scientific Meeting and Exhibition · 2025-09-16

  articleSenior author

  Motivation: Manual segmentation of large MRI datasets is time-consuming. Automated methods could speed up workflows and improve research in epidemiology and clinical settings. Goal(s): We validated the performance of the publicly available MRI deep learning-based TotalSegmentator by comparing it with quality-controlled annotations on 7,064 subjects. Approach: We segmented 55 structures in 28,969 subjects and extracted volume, diameter, and surface area. Bland-Altman plots assessed agreement with manually quality-controlled segmentations in 7,064 cases. Results: Automated segmentation showed high accuracy, though smaller structures like the pancreatic tail posed challenges. Bland-Altman analyses demonstrated strong agreement between automated and quality-controlled segmentations, highlighting the model's scalability and clinical research potential. Impact: This study validates deep learning-based segmentation of the TotalSegmentator model for large-scale MRI analysis (28,969 subjects), showing precise, scalable results. Automated and quality-controlled segmentations demonstrate strong agreement, highlighting its potential to advance research on anatomical structures and health outcomes.

  PublisherDOI

• Imaging-derived biological age across multiple organs links to mortality and aging-related health outcomes

  Research Square · 2025-11-04

  preprintOpen access
  PublisherOA PDFDOI

• Expert-level validation of AI-generated medical text with scalable language models

  Research Square · 2025-07-08 · 1 citations

  preprintOpen access
  PublisherOA PDFDOI

• Biological age assessment in 100,000 whole-body MRI of the German National Cohort (NAKO) and UK Biobank

  Proceedings on CD-ROM - International Society for Magnetic Resonance in Medicine. Scientific Meeting and Exhibition/Proceedings of the International Society for Magnetic Resonance in Medicine, Scientific Meeting and Exhibition · 2025-09-16

  article

  Motivation: MRI provides valuable insights into aging, yet quantifying the true biological age is challenging. Investigations in large, multi-cohort studies can provide comprehensive insights into aging patterns and their influencing factors. Goal(s): This study aims to estimate biological across multiple organs using MRI data from large, multi-cohort studies. Approach: We used ResNet-based models on MRI data from 100,000 participants across two cohorts (UK Biobank and NAKO) to predict organ-specific biological age. Results: The results show that biological age can be estimated for various organs in both cohorts. GradCAM analysis highlights significant regions consistent with known age-related changes. Impact: Our imaging-based multi-organ prediction of biological age from whole-body MRI of 100,000 participants in the UK Biobank and NAKO cohorts provides an important foundation to investigate aging patterns and influencing factors.

  PublisherDOI

• A quest for deep learning MR image reconstruction loss functions in k-space

  Proceedings on CD-ROM - International Society for Magnetic Resonance in Medicine. Scientific Meeting and Exhibition/Proceedings of the International Society for Magnetic Resonance in Medicine, Scientific Meeting and Exhibition · 2025-09-16

  article

  Motivation: Existing loss functions are not optimal for measuring similarity in k-space. Goal(s): To develop an improved k-space loss function that accounts for the k-space characteristics. Approach: We propose a magnitude-phase-energy loss (MPE-loss), which employs separated magnitude and phase losses to improve the accuracy of complex-valued predictions, along with an additional energy term to account for the overall k-space energy. Results: The proposed MPE-loss provides a symmetric loss landscape for each complex k-space sample, creating a quasi-convex loss that facilitates convergence under variations in image brightness and contrast. Experiments demonstrate that MPE-loss outperforms other standard loss functions in k-space reconstruction. Impact: We propose a loss function that is more suitable for k-space, taking into account its characteristics and enhancing the accuracy of k-space regression. This loss function can be applied to any tasks that require calibration of k-space similarity.

  PublisherDOI

• Multi‐Frame Image Registration for Automated Ventricular Function Assessment in Single Breath‐Hold Cine MRI Using Limited Labels

  Magnetic Resonance in Medicine · 2025-10-18 · 4 citations

  articleOpen access

  ABSTRACT Purpose This study aims to develop an automated framework for operator‐independent assessment of cardiac ventricular function from highly accelerated images. Methods We introduce a deep learning framework that generates reliable ventricular volumetric parameters and strain measures from fully sampled and retrospectively accelerated MR images. This method integrates image registration, motion‐compensated reconstruction, and segmentation in a synergetic loop for mutual refinement. The evaluation was performed on an in‐house dataset of healthy and cardiovascular‐diseased subjects. We examined the performance of the underlying tasks, including registration and segmentation, and their impact on derived parameters related to ventricular function. Results The proposed approach demonstrates robustness to undersampling artifacts and requires limited annotation, while still reducing variability and errors for segmentation and registration. This translates to a to increase in Dice similarity compared to existing deep learning methods for left endocardium, left epicardium, and right ventricular delineation. Analysis of the predicted left and right ventricular ejection fraction reveals a strong correlation () with manual measurements. Moreover, the estimated motion and segmentation masks enable consistent radial and circumferential strain measurements across accelerations up to . Conclusion A comprehensive ventricular function analysis can be performed using highly accelerated cine MR data with minimal annotation effort. This multitasking strategy has the potential to enable more accessible cardiac function analysis within a single breath‐hold.

  PublisherDOI

Frequent coauthors

• Konstantin Nikolaou

  129 shared

• Thomas Küstner

  90 shared

• Christian la Fougère

  Universitätsklinikum Tübingen

  69 shared

• Nina F. Schwenzer

  51 shared

• Fabian Bamberg

  University of Freiburg

  47 shared

• Thomas Eigentler

  Freie Universität Berlin

  46 shared

• Christina Pfannenberg

  University Children's Hospital Tübingen

  46 shared

• Holger Schmidt

  University of Tübingen

  42 shared