<b>BiAtria-LGE 100 Dataset</b>

Images and masks are distributed as compressed NIfTI volumes in <code>.nii.gz</code> format.</p><p dir="ltr">The two-cohort structure supports supervised segmentation development, patient-level longitudinal analysis, cross-cohort evaluation, domain generalisation, uncertainty estimation, model calibration, and automated segmentation quality control.</p><h2 dir="ltr"><b>Segmentation Labels</b></h2><table><tr><th><p dir="ltr">Voxel value</p></th><th><p dir="ltr">Anatomical structure</p></th></tr><tr><td><p>0</p></td><td><p dir="ltr">Background</p></td></tr><tr><td><p>1</p></td><td><p dir="ltr">Right atrial wall</p></td></tr><tr><td><p>2</p></td><td><p dir="ltr">Left atrial wall</p></td></tr><tr><td><p>3</p></td><td><p dir="ltr">Right atrial cavity</p></td></tr><tr><td><p>4</p></td><td><p dir="ltr">Left atrial cavity</p></td></tr></table><h2 dir="ltr"><b>Dataset Composition</b></h2><h3 dir="ltr">Utah Cohort</h3><p dir="ltr">The Utah cohort contains <b>91 paired LGE-MRI images and segmentation masks from 38 de-identified patients</b>.</p><p dir="ltr">Several patients have imaging at more than one timepoint, enabling longitudinal analysis.

Available acquisition timepoints include:</p><ul><li>Pre-treatment</li><li>3 months</li><li>4 months</li><li>6 months</li><li>7 months</li><li>8 months</li><li>9 months</li><li>11 months</li></ul><p dir="ltr">The supplied dataset split is defined at the patient level.</p><h4 dir="ltr">Utah Training Partition</h4><ul><li><b>Patients:</b> <code>patient_ID_005</code> to <code>patient_ID_035</code></li><li><b>Number of patients:</b> 31</li><li><b>Number of image-mask pairs:</b> 73</li></ul><h4 dir="ltr">Utah Test Partition</h4><ul><li><b>Patients:</b> <code>patient_ID_001</code> to <code>patient_ID_004</code></li><li><b>Patients:</b> <code>patient_ID_036</code> to <code>patient_ID_038</code></li><li><b>Number of patients:</b> 7</li><li><b>Number of image-mask pairs:</b> 18</li></ul><p dir="ltr">The patient-level partition prevents images from the same Utah patient from appearing in both the training and test datasets.</p><h3 dir="ltr">Waikato Cohort</h3><p dir="ltr">The Waikato cohort contains <b>nine paired LGE-MRI images and segmentation masks</b>.</p><p dir="ltr">The Waikato cohort may be used for external or cross-cohort evaluation.

Researchers should clearly report whether the Waikato cohort was used for training, model selection, calibration, adaptation, validation, or final external testing.</p><h2 dir="ltr"><b>Intended Uses</b></h2><p dir="ltr">BiAtria-LGE 100 is intended for non-clinical research involving:</p><ul><li>Supervised multi-class bi-atrial segmentation</li><li>Right and left atrial wall segmentation</li><li>Right and left atrial cavity segmentation</li><li>Two-dimensional and three-dimensional segmentation methods</li><li>Cross-cohort robustness evaluation</li><li>External validation</li><li>Domain adaptation and domain generalisation</li><li>Transfer learning</li><li>Longitudinal cardiac imaging analysis</li><li>Uncertainty quantification</li><li>Model calibration</li><li>Automated segmentation quality control</li><li>Segmentation failure detection</li><li>Out-of-distribution detection</li><li>Data-efficient and few-shot learning</li><li>Self-supervised representation learning</li><li>Benchmarking of cardiac image segmentation methods</li></ul><h2 dir="ltr"><b>Potential Future Uses</b></h2><h3 dir="ltr">Cross-Cohort Evaluation</h3><p dir="ltr">Models may be trained using the Utah development data and evaluated using the Waikato cohort to investigate how segmentation performance changes across acquisition cohorts.</p><p dir="ltr">Researchers must clearly report whether Waikato images were used during:</p><ul><li>Model training</li><li>Hyperparameter selection</li><li>Model selection</li><li>Preprocessing optimisation</li><li>Threshold selection</li><li>Probability calibration</li><li>Domain adaptation</li><li>Final external evaluation</li></ul><h3 dir="ltr">Longitudinal Segmentation</h3><p dir="ltr">The repeated Utah examinations may support studies of anatomical changes between pre-treatment and post-treatment timepoints.</p><p dir="ltr">Longitudinal studies should account for repeated observations from the same patient.

Scans from the same patient must not be divided between training and evaluation partitions, as this may introduce patient-level data leakage.</p><h3 dir="ltr">Domain Adaptation and Generalisation</h3><p dir="ltr">The two cohorts may support the evaluation of techniques designed to reduce segmentation performance degradation caused by differences in:</p><ul><li>Imaging centres</li><li>Acquisition protocols</li><li>Image reconstruction methods</li><li>Scanner characteristics</li><li>Image intensity distributions</li><li>Patient populations</li></ul><h3 dir="ltr">Uncertainty-Aware Segmentation</h3><p dir="ltr">The dataset may be used to develop methods that estimate prediction uncertainty at the following levels:</p><ul><li>Voxel level</li><li>Slice level</li><li>Anatomical structure level</li><li>Whole-volume level</li><li>Patient level</li></ul><p dir="ltr">Uncertainty estimates may help identify difficult cases or predictions that require manual review.</p><h3 dir="ltr">Automated Quality Control</h3><p dir="ltr">The segmentation masks and image volumes may support the development of methods for detecting:</p><ul><li>Missing anatomical classes</li><li>Anatomically implausible segmentations</li><li>Disconnected components</li><li>Excessive false-positive regions</li><li>Wall-cavity label reversal</li><li>Segmentation leakage</li><li>Empty masks</li><li>Abnormally small or large atrial volumes</li><li>Low-confidence predictions</li></ul><h3 dir="ltr">Few-Shot External Adaptation</h3><p dir="ltr">The smaller Waikato cohort may support research involving:</p><ul><li>Few-shot fine-tuning</li><li>Active learning</li><li>Parameter-efficient adaptation</li><li>Test-time adaptation</li><li>Semi-supervised learning</li><li>Unsupervised domain adaptation</li></ul><h3 dir="ltr">Model Calibration</h3><p dir="ltr">Researchers may investigate whether predicted probabilities correspond to actual segmentation correctness and whether model calibration changes between the Utah and Waikato cohorts.</p><p dir="ltr">Calibration may be evaluated at the voxel, anatomical structure, case, or cohort level.</p><h2 dir="ltr"><b>Notes for Researchers</b></h2><p dir="ltr">Researchers are encouraged to use the patient-level training and test split defined in the accompanying <code>patient_image_mapping.csv</code> file.</p><p dir="ltr">Any alternative split must also be defined at the patient level.

All scans belonging to the same patient should remain within a single training, validation, or test partition.</p><p dir="ltr">Researchers should report:</p><ul><li>The exact training, validation, and test partitions</li><li>Whether the recommended split was used</li><li>Whether any cases were excluded</li><li>Whether Waikato data were used during model development</li><li>Cohort-specific results for Utah and Waikato</li><li>Class-wise segmentation results</li><li>Preprocessing and postprocessing procedures</li><li>Model calibration and uncertainty methods, where applicable</li><li>Representative segmentation failures</li></ul><h2 dir="ltr"><b>

Citation

Requirements</b></h2><p dir="ltr">Researchers using this dataset must cite both:</p><ol><li>The <b>BiAtria-LGE 100 Figshare dataset record</b>, using the citation generated by Figshare.</li><li>The associated publication listed under "Related Materials"</li></ol><p><br></p>