Endometrial cancer (EC) is one of the most common malignant tumors of the female reproductive system, primarily occurring in postmenopausal women and closely linked to obesity, aging, and metabolic conditions. Although symptoms like postmenopausal bleeding often prompt early evaluation, the clinical picture frequently overlaps with benign conditions like endometrial hyperplasia.
While MRI provides excellent soft tissue contrast and is widely used in EC evaluation, traditional MRI interpretation depends on clinical experience and is prone to subjectivity. MRI often fails to clearly delineate the depth of myometrial invasion and lymph node involvement - two factors critical for surgical and treatment planning.
AI-based deep learning offers the possibility of consistently extracting complex quantitative features from MRI data that go beyond what human visual interpretation captures, potentially improving both diagnostic accuracy (distinguishing high-risk from low-risk EC) and prognostic prediction (identifying patients likely to experience postoperative recurrence).
The core architecture was ResNet-101, a 101-layer residual convolutional neural network that has proven highly effective for medical image analysis. ResNet's residual connections allow deep networks to be trained without degradation, enabling very deep feature learning from MRI images.
The researchers enhanced standard ResNet-101 by incorporating two types of attention mechanisms: a spatial attention module, which emphasizes which regions of the image are most important, and a channel attention module, which emphasizes which feature maps (types of learned information) carry the most diagnostic relevance. Together, these allow the model to focus on the most informative portions of each MRI image.
The dataset comprised MRI scans from 210 EC patients, split into a training set of 140 cases and a validation set of 70 cases. The validation set included 45 low-risk and 25 high-risk patients. The improved model was compared head-to-head against three alternatives: standard ResNet-101, spatial-attention-only (SA-ResNet-101), and channel-attention-only (CA-ResNet-101).
For the task of diagnosing high-risk endometrial cancer, the combined spatial-channel attention model achieved an AUC of 0.918. This was dramatically higher than the comparators: standard ResNet-101 scored 0.613, SA-ResNet-101 scored 0.760, and CA-ResNet-101 scored 0.758. The difference was statistically significant (p less than 0.05) across all performance metrics.
The full set of diagnostic metrics - accuracy (AC), precision (PR), recall (RE), and F1 score - were all visibly higher for the proposed model. The F1 score is the harmonic mean of precision and recall, providing a balanced measure of performance particularly relevant when the positive class (high-risk cancer) is rarer than the negative class.
These results demonstrate that incorporating both spatial and channel attention is synergistic: neither attention mechanism alone achieves anywhere near the performance of their combination, suggesting the two mechanisms capture complementary aspects of the discriminative information present in EC MRI images.
For postoperative recurrence prediction, the model demonstrated even stronger performance, achieving an AUC of 0.926. Among the 70 validation patients, 13 experienced postoperative recurrence and 57 did not, creating an imbalanced prediction task where correctly identifying the recurrent cases is clinically most important.
Again, the proposed model substantially outperformed alternatives: standard ResNet-101 scored 0.620, SA-ResNet-101 scored 0.729, and CA-ResNet-101 scored 0.767 for recurrence prediction. The improvement was statistically significant (p less than 0.05) across all metrics.
The superior recurrence prediction performance is particularly clinically valuable because identifying which patients will relapse after surgery could guide decisions about adjuvant therapy - additional treatment given after surgery to reduce recurrence risk. Patients flagged as high-recurrence-risk by the model could receive more intensive follow-up or additional treatment consideration.
The combined AUC values above 0.90 for both high-risk diagnosis and recurrence prediction from routine preoperative MRI suggest the model could meaningfully change clinical decision-making. A high-risk classification could trigger more extensive lymph node dissection at surgery, while a high-recurrence-risk prediction could initiate adjuvant chemotherapy or radiation planning immediately after surgery.
The study's use of a retrospective dataset of 210 patients is a limitation - prospective validation in an independent cohort is needed before clinical deployment. However, the consistent and large performance gap over comparator models provides strong evidence that the dual-attention architecture captures genuinely predictive features from EC MRI.
The approach exemplifies a broader trend in AI-assisted radiology: rather than replacing radiologists, AI tools function as decision support systems that enhance radiologist accuracy and consistency, particularly in the quantitative assessment of features like invasion depth and tissue characteristics that are difficult to measure reliably by eye alone.