MRI plays a critical role in planning treatment for endometrial cancer because it reveals how deeply the tumor has invaded the uterine wall, whether it has spread to the cervix or lymph nodes, and other features that determine surgical approach. In recent years, radiomics -- extracting hundreds of quantitative image features from precisely outlined tumor regions -- has emerged as a powerful tool for predicting recurrence risk and guiding personalized treatment decisions. However, every radiomics study depends on tumor segmentation, the process of drawing the exact boundary of the cancer on each image slice.
Manual segmentation by a radiologist is time-consuming, subjective, and prone to variability between different observers. For endometrial cancer specifically, the challenge is amplified because the normal endometrium can look similar to early cancer on some MRI sequences, requiring the radiologist to integrate information across multiple image types simultaneously. This creates a bottleneck that limits how many cases can be analyzed and introduces measurement inconsistency. Automating this step with a neural network would enable high-throughput, reproducible radiomics analysis at scale.
The study used 200 patients with pathologically confirmed endometrial cancer who underwent preoperative MRI at Kyoto University Hospital between 2004 and 2017. 180 patients were used for training and 20 for independent testing. The neural network was a modified U-net -- a fully convolutional architecture widely used for medical image segmentation. The modification introduced five processing blocks with batch normalization and Leaky ReLU activations, and used Dice loss as the training objective. Dice loss directly optimizes the overlap between predicted and true tumor areas, making it better suited for segmentation tasks than standard cross-entropy loss, particularly when tumor pixels are rare compared to background pixels.
Three different input configurations were tested: single-sequence images from T2WI, DWI, or ADC map alone, and a multi-sequence input combining all three as separate channels (analogous to RGB channels in color photography). The study also systematically tested the effect of batch size -- the number of image slices processed per training step -- ranging from 11 to 176, which required a high-memory 96-GB GPU to achieve the largest values. Data augmentation techniques (mix-up and random image cropping and patching) were applied to reduce overfitting on the limited training set.
On the independent 20-patient test set, the multi-sequence model (T2WI plus DWI plus ADC map) achieved the best performance: mean Dice similarity coefficient (DSC) of 0.806, sensitivity of 0.816, and positive predictive value (PPV) of 0.834. A DSC of 0.806 means that 80.6% of pixels are correctly identified as belonging to the tumor, with a median DSC of 0.87 across patients. The T2WI-only model was close (DSC 0.798) but statistically similar; the DWI-only and ADC-only models were significantly worse (DSC 0.679 and 0.557 respectively, both p less than 0.001 vs. multi-sequence).
The performance advantage of the multi-sequence model makes biological sense: while endometrial cancer shows high signal on DWI due to dense cell packing, normal endometrium also shows high DWI signal, making DWI alone ambiguous. Radiologists routinely integrate multiple sequences to establish tumor boundaries in clinical practice, and the multi-sequence model replicates this approach. The result also outperformed the only prior published automated EC segmentation study, which used contrast-enhanced T1-weighted imaging alone and achieved median DSC values of 0.84 and 0.77 for two human raters -- compared to the 0.87 achieved here without contrast enhancement.
A key technical finding was the strong effect of batch size on segmentation quality. The Dice loss on the cross-validation set dropped progressively from 0.592 with batch size 11 to 0.231 with batch size 176 -- a dramatic improvement. The explanation is structural: the entire MRI volume of each patient was used as input (all axial slices), but only a small fraction of those slices contain tumor tissue. With a small batch size, it is common for a training step to contain only non-tumor slices, giving the model nothing to learn from. Larger batches statistically guarantee more tumor-containing slices per update, providing a richer learning signal each step.
The conventional U-net architecture (which starts with 64 filters in the first layer) could not physically run at batch sizes above 44 because of GPU memory limitations, and its performance plateaued at Dice loss 0.267. The modified architecture was designed with fewer filters in early layers, enabling the very large batch sizes that drove the performance improvement. This architectural choice -- sacrificing some model complexity in the early layers to enable larger batch training -- proved to be the right trade-off for this specific segmentation task.
Beyond segmentation accuracy, the study evaluated whether radiomic features extracted from automatically segmented tumors agree with those extracted from manually segmented tumors -- a critical test for clinical usability. Agreement was measured using the intraclass correlation coefficient (ICC), where values above 0.75 indicate good reliability and above 0.9 indicate excellent reliability. First-order features (statistics describing the distribution of pixel intensities: mean, median, entropy, etc.) showed good-to-excellent reliability with a median ICC of 0.86. Shape-based features (describing the 3D geometry of the tumor: volume, surface area, elongation, flatness) showed the highest reliability with a median ICC of 0.96.
Higher-order texture features showed more variable results. Features from the gray-level size zone matrix (GLSZM) and gray-level dependence matrix (GLDM) showed good reliability (median ICC 0.93 and 0.87 respectively), while features from the gray-level co-occurrence matrix (GLCM) showed only moderate reliability (median ICC 0.57). This variability reflects how sensitive texture features are to small differences in segmentation boundaries -- even a few misidentified pixels at the tumor edge can change these calculations. The results suggest that first-order and shape features from automated segmentation can be reliably used in downstream radiomics analysis, while some texture features should be interpreted with caution.
The ability to automatically segment endometrial cancer tumors with DSC 0.806 and extract reliable radiomic features has direct practical consequences. Building large radiomics datasets for endometrial cancer currently requires radiologists to manually outline tumors in hundreds of patients -- a process taking many hours per study. Automation would allow this to be done in seconds per patient, enabling the large-scale, reproducible datasets needed to develop clinically validated prediction models for recurrence risk, lymph node involvement, and treatment response. The model's ability to achieve these results without contrast-enhanced imaging is particularly valuable: some patients cannot receive contrast agents due to kidney problems or allergies, so a non-contrast model is more broadly applicable.
The study's limitations include its single-center, single-vendor design -- all MRI scans were acquired on Siemens systems at one hospital using one protocol. Whether the model generalizes to different scanner brands, field strengths, or imaging protocols at other institutions is unknown and requires multicenter validation. Additionally, the 20-patient test set is small, meaning the performance estimates carry uncertainty. The authors identify multicenter validation and the potential addition of contrast-enhanced imaging as the key next steps toward clinical deployment.