MRI (magnetic resonance imaging) is the gold standard imaging modality for preoperative evaluation of endometrial cancer. It provides detailed soft-tissue contrast that helps surgeons and radiation oncologists understand the size of the tumor, how deeply it has invaded the uterine muscle (myometrium), and whether it has spread to adjacent structures. This information is critical for surgical planning and for determining which patients need additional treatment after surgery.
Tumor segmentation is the process of precisely delineating the tumor boundaries on medical images - essentially drawing an outline around the tumor in each imaging slice. Accurate segmentation enables quantitative measurements of tumor volume, shape, and heterogeneity that are increasingly used in treatment planning, response monitoring, and radiomics research (the extraction of large numbers of quantitative features from medical images for predictive modeling).
Currently, tumor segmentation is performed manually by radiologists or oncologists, a process that is time-consuming, requires significant expertise, and is subject to variability between different clinicians (inter-rater variability). An automated, reliable segmentation system would save significant clinical time and improve consistency of measurements - both important for large-scale research studies and for routine clinical workflow.
The study included 139 endometrial cancer patients who underwent preoperative pelvic MRI using a T1-weighted VIBE sequence (Volume Interpolated Breath-hold Examination), a technique that provides high-resolution three-dimensional images of the pelvis. The dataset was divided into training, validation, and test cohorts, with 34 patients forming the final test cohort on which model performance was evaluated.
The deep learning architecture used was a 3D UNet - a convolutional neural network designed specifically for medical image segmentation. Unlike 2D networks that process images slice by slice, the 3D UNet processes volumetric image data simultaneously, allowing it to exploit spatial context in all three dimensions. The U-shaped architecture has encoding pathways that extract features at multiple scales and decoding pathways that reconstruct spatial detail at the original image resolution.
Two expert clinicians (referred to as Rater 1 and Rater 2) independently manually segmented all tumors in the test cohort. Their segmentations served both as training labels and as the gold standard for evaluating algorithm performance. Crucially, comparing inter-rater disagreement (R1 versus R2) to algorithm-rater disagreement (ML versus R1, ML versus R2) reveals whether the algorithm performs within the natural range of human expert variation.
The primary performance metric used was Dice Similarity Coefficient (DSC) - a measure of the overlap between two segmentations ranging from 0 (no overlap) to 1 (perfect overlap). The median DSC between the two human raters (R1 vs R2) was 0.89, representing the gold standard of expert agreement. The algorithm's median DSC against Rater 1 was 0.84, and against Rater 2 was 0.77.
A DSC of 0.84 is considered clinically acceptable for medical image segmentation tasks. The difference between the algorithm's agreement with Rater 1 (DSC 0.84) and the agreement between the two human raters (DSC 0.89) was not statistically significant (p = 0.08), indicating that the algorithm performs within the range of inter-expert variability. In practical terms, the algorithm's segmentations were nearly as consistent with expert opinion as the experts were with each other.
Additional metrics including Hausdorff Distance (HD) - which measures the maximum distance between the borders of two segmentations - and Average Surface Distance (ASD) - which measures the average difference in boundary location - showed similar patterns. No significant difference in tumor volume measurements was found between raters and the machine learning model (p = 0.28), confirming that the algorithm produces volumetrically accurate segmentations.
A key design choice in this study was the use of three-dimensional volumetric segmentation rather than slice-by-slice (2D) analysis. Uterine tumors are complex three-dimensional structures, and their boundaries in a given 2D slice depend on context from neighboring slices. A 3D network can directly capture these volumetric relationships, potentially producing more anatomically coherent segmentations than 2D approaches that must stitch together independently processed slices.
The VIBE MRI sequence used in this study is particularly well-suited for volumetric analysis because it is acquired as a true 3D dataset with isotropic (equal in all directions) or near-isotropic voxel sizes - the three-dimensional equivalent of pixels. This means the same spatial resolution is available in all three dimensions, making the 3D UNet's volumetric learning approach maximally effective.
One practical challenge for 3D networks is computational demand: processing entire volumes requires more memory and computing power than processing 2D slices. The researchers addressed this through data preprocessing steps including image normalization and potentially patch-based training (training on smaller 3D sub-volumes rather than the full scan), which is standard practice for medical volume segmentation with deep learning on current hardware.
Automated tumor segmentation has immediate practical value for radiomics research. Radiomics involves extracting hundreds to thousands of quantitative features from medical images (such as texture, shape, and intensity statistics within the tumor) to predict outcomes like lymph node metastasis, myometrial invasion depth, and treatment response. These analyses require precise, reproducible segmentations - exactly what an automated system can provide at scale.
In radiation therapy planning, accurate tumor delineation determines which tissue receives the highest radiation dose while neighboring healthy structures are spared. The time required for manual segmentation is a bottleneck in radiotherapy workflows, particularly for complex pelvic cases. An automated system achieving expert-level accuracy would reduce this burden significantly, potentially allowing radiation oncologists to see more patients or devote more review time to complex cases.
The study also highlights the importance of evaluating AI systems against inter-rater variability rather than against a single ground truth annotation. Because even expert clinicians disagree on segmentation boundaries, any automated system will inevitably fall short of perfect agreement with any single annotator. Framing performance relative to human-human disagreement provides a more realistic and clinically meaningful benchmark for assessing whether an algorithm is ready for clinical use.
This study demonstrates that a 3D deep learning model can automate endometrial tumor segmentation to a standard that falls within the range of expert-to-expert agreement. For a 34-patient test cohort, the system produced segmentations with tumor volume measurements statistically indistinguishable from manual expert measurements, and boundary accuracy (DSC 0.84) approaching the inter-expert standard (DSC 0.89).
The clinical implications extend beyond simply saving radiologist time. Standardized, reproducible automated segmentations could enable multicenter collaboration on large-scale radiomics studies that would be impractical with manual annotation, potentially accelerating the discovery of imaging biomarkers for endometrial cancer staging and prognosis.
Limitations include the relatively small test cohort of 34 patients and the single-institution nature of the dataset. External validation on MRI scans from other centers, acquired with different scanners and protocols, will be necessary to confirm that the algorithm generalizes reliably. Integration into clinical picture archiving and communication systems (PACS) as a pre-segmentation tool for radiologist review represents a near-term translational goal.