The most critical staging decision in early endometrial cancer is distinguishing FIGO Stage IA (tumor confined to the inner half of the uterine muscle) from Stage IB (tumor invading the outer half). This distinction changes both the surgical approach and the recommendation for adjuvant treatment. MRI is the imaging modality of choice for this assessment because of its superior soft tissue contrast, but accurate staging remains difficult even for experienced radiologists.
Conventional MRI staging is performed by radiologists who visually assess the depth of myometrial invasion from two-dimensional image slices. This process is subjective, time-consuming, and subject to inter-reader variability. Deep learning approaches that automatically classify MRI scans could provide consistent, objective staging while reducing radiologist workload.
Most prior deep learning approaches to medical image classification use single imaging planes (typically axial/transverse). However, endometrial cancer staging benefits from simultaneous assessment of the tumor from multiple angles - the same principle that guides radiologists to review coronal, sagittal, and transverse views together before reaching a conclusion. This study proposes a deep learning architecture that explicitly fuses information from all three planes.
The study uses the Swin Transformer as its backbone architecture. Unlike traditional CNNs, transformers use self-attention mechanisms to capture long-range dependencies between distant image regions - a capability particularly valuable for assessing tumor extent across an entire MRI slice. The Swin Transformer specifically uses Shifted Window Multi-Head Self-Attention (SW-MSA), which applies attention within local windows that shift between transformer layers, enabling efficient computation while maintaining global context.
Two additional components are incorporated: a Squeeze-and-Excitation (SE) module, which recalibrates feature channel importance through global average pooling and learned channel weights, and a multi-axis attention module that processes spatial attention along different axes of the image independently before combining results. Together these modifications enhance the model's sensitivity to subtle spatial patterns that indicate myometrial invasion depth.
The dataset comprised 122 patients: 68 with Stage IA and 54 with Stage IB disease, each with T2-weighted MRI (T2WI) scans in all three standard planes (sagittal, coronal, and transverse). This multi-plane dataset enables both single-plane and multi-plane fusion experiments, allowing direct comparison of the benefit gained from combining views.
Single-plane classification using the proposed model achieved strong results in all three planes: sagittal accuracy 0.988, coronal accuracy 0.96, transverse accuracy 0.94. The sagittal plane performed best, which is consistent with clinical practice where the sagittal view is typically considered most informative for assessing anterior-posterior uterine anatomy and myometrial invasion depth.
Remarkably, multi-plane fusion achieved an accuracy of 1.0 (100%) on the test set - perfect classification of all Stage IA and IB cases when all three MRI planes were combined. While this result should be interpreted with caution given the modest dataset size, it strongly demonstrates that fusion of complementary information from multiple viewing angles provides benefits beyond any single plane.
The proposed architecture outperformed multiple comparison methods including ViT-B/16 (the standard vision transformer), ResNet50 (a leading CNN), and a traditional SVM classifier using manually engineered HOG features. This demonstrates gains from both the transformer architecture and the multi-plane fusion strategy over both deep learning and classical machine learning baselines.
Multi-plane fusion in this study works by processing each MRI plane (sagittal, coronal, transverse) independently through the same Swin Transformer model, producing three separate feature vectors. These features are then combined - either through concatenation or learned fusion weights - before the final classification layer makes a staging prediction.
The rationale for this approach is that each plane captures different anatomical perspectives on the same tumor. A shallow invasion that appears ambiguous in the transverse plane may be clearly visible as confined to the inner myometrium in the sagittal plane. By presenting all three perspectives simultaneously, the fusion model can resolve ambiguities that any single view cannot.
The Swin Transformer's windowed self-attention is particularly well-suited to this task because myometrial invasion is a spatial relationship - the proximity of the tumor margin to the uterine serosa - that requires attention to both local texture and global position within the organ. The shifted window design allows features from both scales to interact, which is difficult to achieve with standard CNN architectures using fixed receptive fields.
Accurate preoperative staging directly affects surgical planning. Stage IB disease has a higher risk of lymph node involvement than Stage IA, and many guidelines recommend sentinel lymph node mapping or lymphadenectomy for Stage IB but not Stage IA patients. Misclassification in either direction carries consequences: understaging may lead to inadequate lymph node assessment, while overstaging may result in unnecessary procedures.
A deep learning tool achieving near-perfect staging accuracy could eventually reduce reliance on surgeon experience and institutional expertise for this critical preoperative decision. This is particularly important for patients treated at centers with limited subspecialty gynecological oncology expertise, where preoperative MRI staging accuracy may be lower than at academic centers.
The non-invasive nature of the classification - based entirely on standard T2-weighted MRI sequences already obtained in routine preoperative workup - means no additional testing would be required to apply this tool. Implementation would require only software integration into the radiology reporting workflow and validation in diverse clinical populations.
The principal limitation of this study is the small dataset of 122 patients from a single institution. The perfect 100% multi-plane fusion accuracy, while exciting, should be interpreted cautiously because small test sets can produce extreme performance metrics through chance. A test set of even 30-40 patients means a single misclassification would drop accuracy by 2-3 percentage points, so the reported 100% reflects zero errors in a limited sample.
Validation in larger multicenter datasets from multiple MRI scanner manufacturers, magnetic field strengths (1.5T versus 3T), and imaging protocols is essential. MRI images vary substantially across centers due to equipment and protocol differences, and a model trained at one site may perform significantly worse at another without domain adaptation or retraining.
Despite these limitations, this work represents a technically sophisticated and clinically motivated approach to endometrial cancer staging. The multi-plane fusion strategy is conceptually sound and the achieved accuracy levels - even at single-plane level - suggest genuine potential. Larger validation studies should prioritize this architecture as a serious candidate for clinical-grade AI-assisted staging.