Accurate segmentation of uterine lesions on MRI scans is essential for staging endometrial cancer and planning surgery. Traditionally, radiologists must manually outline each lesion on dozens or hundreds of image slices, a process that is slow, labor-intensive, and subject to variability between clinicians.
In June 2023, the International Federation of Gynecology and Obstetrics (FIGO) updated its staging system for endometrial cancer to make it more precise, which placed even higher demands on image analysis accuracy. Automated segmentation tools are urgently needed to meet these demands.
Most high-performing deep learning models require pixel-level labels, meaning expert radiologists must carefully outline every lesion boundary on every image slice used for training. This process is extremely time-consuming and expensive, creating a major bottleneck in developing AI tools for medical imaging.
This study aimed to develop a model that uses only image-level annotations, where a radiologist simply marks whether a given scan contains a specific lesion type rather than outlining exact boundaries. This dramatically reduces the annotation burden while still producing useful segmentation results.
The model was trained to segment four types of uterine lesions: endometrial cancer (EC), uterine leiomyoma (benign muscle tumors), endometrial polyps (growths on the uterine lining), and atypical hyperplasia of the endometrium (abnormal cell overgrowth that can progress to cancer).
In addition to lesions, the model also needed to distinguish three types of normal tissue: the uterus body, the cervix, and the endometrium (inner lining of the uterus). This multi-class segmentation task is considerably more complex than identifying a single lesion type.
The dataset came from 316 patients at Dalian Women and Children's Medical Group in China. All scans used T2-weighted MRI sequences, which provide good contrast between soft tissues. The full dataset contained 85,730 image slices. Patients were split 60/20/20 for training, validation, and testing.
Because endometrial polyps and atypical hyperplasia cases were relatively rare in the dataset, the researchers used data augmentation, randomly scaling, flipping, and rotating existing images to create more training examples and balance the proportion of each lesion type.
The proposed system uses a two-stage, four-module pipeline. The first stage generates rough maps of where lesions are likely located (called Class Activation Maps or CAMs), and the second stage refines these maps into precise segmentation masks.
Stage one begins with the Pixel Correlation Module (PCM), which uses a self-attention mechanism to refine initial CAMs by considering how similar neighboring pixels are to each other. A standard ResNet-50 neural network serves as the backbone for feature extraction.
The Class Re-Activation Map (ReCAM) module then converts the multi-lesion detection task into separate single-lesion tasks, preventing different lesion types from interfering with each other. The Inter-pixel Relation Network (IRNet) further sharpens boundaries by analyzing spatial relationships between pixels and generating high-quality pseudo segmentation masks.
In stage two, these pseudo masks are used as training labels for DeepLab v3+, a powerful fully-supervised segmentation network. This final module uses atrous convolutions and an encoder-decoder architecture to produce detailed, multi-scale segmentations from the refined pseudo-label training data.
The final model achieved an average Dice Similarity Coefficient (DSC) of 83.5% on the test set, meaning the model's predicted boundaries overlapped with expert-drawn boundaries about 83.5% of the time. Higher DSC indicates better agreement with the gold standard.
Among specific lesion types, uterine leiomyoma was segmented most accurately (DSC 82.4%), while atypical hyperplasia of endometrium was the most challenging, with the lowest DSC at 71.3%. This likely reflects the subtler visual characteristics and smaller size of hyperplastic lesions.
The model's Hausdorff Distance (HD = 29.3 mm) and Average Symmetric Surface Distance (ASSD = 8.83 mm) both indicate that predicted boundaries were reasonably close to actual lesion edges. Lower values for both metrics indicate better boundary accuracy.
An ablation study showed each module contributed to the final performance. The baseline CAM alone achieved 64.9% DSC; adding PCM improved it to 70.2%, adding ReCAM alone gave 70.9%, and adding IRNet made the largest single contribution, boosting DSC by 10.3 percentage points.
Segmenting uterine lesions is inherently difficult because lesion boundaries are often not clearly defined. Adhesion between a lesion and surrounding normal tissue creates ambiguous borders where even expert radiologists disagree, making a definitive ground truth difficult to establish.
Uterine lesions tend to have irregular shapes that do not conform to standard geometric models, and their size relative to the full image is often small. This imbalance between lesion pixels and background pixels poses a challenge for training AI models that might default to ignoring rare lesion regions.
The researchers addressed the small-lesion problem by cropping a 470x435 pixel region of interest (ROI) centered on the uterus from each larger image. This focused the model on relevant anatomy and improved the proportion of lesion-containing pixels in training batches.
Compared to prior state-of-the-art weakly supervised methods designed for natural images, the proposed model demonstrated superior or equivalent performance when all methods were retrained on the same uterine MRI dataset, confirming that the approach is well-suited to this specialized medical imaging domain.
Automated segmentation tools could significantly reduce radiologist workload in gynecological imaging. Instead of outlining every lesion by hand, clinicians could use AI-generated contours as a starting point for review and correction, saving time while maintaining accuracy.
For endometrial cancer specifically, accurate segmentation is critical for determining surgical staging and selecting between conservative versus aggressive surgical approaches. Errors in tumor boundary delineation can directly affect treatment decisions and patient outcomes.
The model's ability to detect multiple lesion types simultaneously also has practical advantages, as many patients present with more than one uterine pathology at the same time. A comprehensive segmentation tool reduces the need for separate specialized models for each condition.
The study's main limitations include its single-center dataset from one hospital in China and the relatively modest number of patients for some lesion types. Future work should validate the model in diverse clinical settings, with larger datasets and prospective evaluation in real-world practice.