Endometrial cancer prevalence: Endometrial cancer (EC) accounts for 15% to 20% of all malignant tumors in the female reproductive system and is most common in women aged 50 to 60 years. Accurate staging of the tumor is essential for determining the appropriate treatment plan.
Limitations of other imaging modalities: Conventional CT struggles to distinguish stage II, III, and IV EC and is limited to evaluating intrauterine invasion. Ultrasound is often insufficiently accurate for staging. MRI, by contrast, offers superior soft tissue resolution, multiplanar imaging, and high accuracy - making it the preferred preoperative staging tool.
Role of deep learning: MRI image analysis for EC traditionally relies on qualitative assessment by radiologists, limiting reproducibility and quantitative precision. Deep learning can improve feature extraction and classification efficiency while providing quantitative measurements from MRI data.
Quantitative MRI challenge: Achieving quantitative, timing-based MRI analysis requires clean training data and robust networks. Patient motion artifacts during MRI acquisition remain a major obstacle, and single-scan quantitative imaging methods need continued development.
Patient cohort: Eighty patients with EC were enrolled, aged 42 to 70 years (average 51.67 years). Disease subtypes included 32 cases of endometrioid adenocarcinoma, 9 cases of low differentiation, 10 cases of high differentiation, 6 cases of serous adenocarcinoma, and 2 cases of carcinosarcoma.
MRI acquisition: A 1.5 Tesla MRI machine was used with both coronal and sagittal T1-weighted imaging (T1WI) and T2-weighted imaging (T2WI), plus diffusion-weighted imaging (DWI) and dynamic contrast-enhanced (LAVA) sequences. Gadolinium-based contrast (Gd-DTPA) was administered intravenously for multiphase dynamic scanning.
Staging framework: Two experienced radiologists used a double-blind method to evaluate endometrial thickness, cervical canal width, fibrous matrix ring integrity, signal characteristics, and lymph node status. They applied the 2009 International Federation of Gynecology and Obstetrics (FIGO) staging criteria.
Staging categories: Stage I is limited to the uterus (Ia: less than 50% myometrial invasion; Ib: more than 50%). Stage II involves cervical involvement. Stages III and IV indicate extrauterine spread to organs, lymph nodes, or distant metastasis.
Three network approaches compared: The study evaluated three convolutional network modes for MRI image classification: a shallow CNN network, a standard ResNet network, and an optimized ResNet network. All were trained and tested on a dataset of 867 training images and 436 test images.
ResNet design principle: ResNet avoids vanishing gradient problems through residual (skip) connections. Each layer learns a residual function relative to its input rather than learning the full mapping from scratch, which allows networks to be trained at much greater depth without accuracy degradation.
Optimized ResNet modifications: The authors improved upon standard ResNet by setting input image size to 100x100 pixels, changing step sizes to 1 in the first bottleneck residual block, adding the Adam optimizer, and using ReLU activation functions throughout. These changes improved model accuracy at greater network depth.
Loss function and training: A distance-based loss function minimizes the Frobenius norm between the reconstructed image and the ground-truth label. Gradient descent iteratively updates network weights and thresholds to minimize this loss, with training proceeding until error requirements are satisfied.
Network comparison results: The optimized ResNet network demonstrated the best performance among the three models, with the standard ResNet second and the shallow CNN last. The optimized ResNet processing time was 0.48 seconds per image compared to 0.14 seconds for the shallow CNN, reflecting its greater model complexity.
Diagnostic accuracy metrics: For MRI T2-weighted imaging combined with enhanced scanning, the optimized approach achieved 88.75% accuracy, 95% specificity, 87.5% sensitivity, 93.75% negative predictive value, and 86.25% positive predictive value.
Staging accuracy: Of 80 EC patients, MRI correctly diagnosed 72 as Stage I (90%). The detailed breakdown showed 56 true Stage Ia shallow muscle layer infiltration cases correctly identified, plus 15 Stage II cervical infiltration cases also diagnosed as Stage Ia - indicating a tendency to underestimate advanced staging in some cases.
MRI appearance characteristics: The study confirmed that T2WI shows endometrial thickening in EC patients with slightly high signal intensity - lower than healthy endometrium but higher than the muscular layer. Stage I tumors typically showed widening of the high-signal area in the uterine body on sagittal T2WI.
MRI strengths confirmed: The study confirms prior literature showing MRI accuracy of 92% to 97% for deep myometrial infiltration and 69% to 74% for shallow infiltration. MRI is particularly effective for Stage I diagnosis and for evaluating cervical infiltration and lymph node metastasis.
Comparison with prior work: Tsuyoshi et al. achieved 100% sensitivity, 96.9% specificity, and 97.0% accuracy using PET/MRI for lymph node metastasis detection. Bi et al. demonstrated that combining T2-weighted imaging with dynamic contrast-enhanced MRI and diffusion-weighted imaging yields the highest diagnostic accuracy with high specificity.
Advantages of the deep learning approach: Integrating the optimized ResNet network into MRI imaging effectively improves the accuracy of image feature extraction and reduces dependence on individual radiologist expertise. Deep learning enables more reproducible and objective analysis of complex MRI sequences.
Study limitations: The sample size of 80 patients is relatively small. The experimental data cannot completely eliminate subjective factors. Future work should expand to larger cohorts with additional physiological and clinical variables to strengthen the diagnostic model.
Key finding: Applying an optimized ResNet deep learning model to MRI imaging of EC patients significantly improves the efficiency of image feature extraction and classification. The system achieves 88.75% accuracy and 95% specificity with high clinical utility for preoperative staging.
Practical implications: MRI combined with deep learning provides high-resolution, multiplanar imaging with quantitative capability that can assist clinicians in staging decisions. The system is particularly strong for Stage I diagnosis, which represents the majority of EC cases and where early detection has the greatest impact on outcomes.
Future directions: The authors acknowledge the need for larger datasets and plan to incorporate more physiological and clinical data to further refine and expand the diagnostic model. Standardizing next-stage evaluation metrics and eliminating subjective interference are priorities for future validation work.