Organizations: School of Medical Information and Engineering, Southwest Medical University, Luzhou 646000, Sichuan, China · School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu 611731, China · Department of Ultrasound, Affiliated Hospital of Southwest Medical University, Luzhou 646000, Sichuan, China · Department of Functional Examination Unit, Zibo Hospital of Traditional Chinese Medicine, Zibo 255000, China · Key Laboratory of Medical Electrophysiology, Ministry of Education& Medical Electrophysiological Key Laboratory of Sichuan Province, Institute of Cardiovascular Research, Southwest Medical University · Department of Radiology, the First Affiliated Hospital of Chongqing Medical University, Chongqing 400016, China · International Research Center for Complexity Sciences, Hangzhou International Innovation Institute, Beihang University, Hangzhou 311115, China · Institute of Intelligent Chinese Medicine, Chongqing University of Chinese Medicine, Chongqing 402760, China · Department of Cardiology, Affiliated Hospital of Southwest Medical University, Luzhou 646000, Sichuan, China · Basic Medicine Research Innovation Center for Cardiometabolic Diseases, Ministry of Education, Southwest Medical University · Nucleic Acid Medicine of Luzhou Key Laboratory, Southwest Medical University, Luzhou 646000, Sichuan, China
Early detection of myometrial invasion is critical for the staging and life-saving management of endometrial carcinoma (EC), a prevalent global malignancy. Transvaginal ultrasound serves as the primary, accessible screening modality in resource-constrained primary care settings; however, its diagnostic reliability is severely hindered by low tissue contrast, high operator dependence, and a pronounced scarcity of positive pathological samples. Existing artificial intelligence solutions struggle to overcome this severe class imbalance and the subtle imaging features of invasion, particularly under the strict computational limits of primary care clinics. Here we present an automated, highly efficient two-stage deep learning framework that resolves both data and computational bottlenecks in EC screening. To mitigate pathological data scarcity, we develop a structure-guided cross-modal generation network that synthesizes diverse, high-fidelity ultrasound images from unpaired magnetic resonance imaging (MRI) data, strictly preserving clinically essential anatomical junctions. Furthermore, we introduce a lightweight screening network utilizing gradient distillation, which transfers discriminative knowledge from a high-capacity teacher model to dynamically guide sparse attention towards task-critical regions. Evaluated on a large, multicenter cohort of 7,951 participants, our model achieves a sensitivity of 99.5%, a specificity of 97.2%, and an area under the curve of 0.987 at a minimal computational cost (0.289 GFLOPs), substantially outperforming the average diagnostic accuracy of expert sonographers. Our approach demonstrates that combining cross-modal synthetic augmentation with knowledge-driven efficient modeling can democratize expert-level, real-time cancer screening for resource-constrained primary care settings.
Figures & tables
Table 1: Comparison of Image Quality Metrics.
Figure 1: Visual comparison of MRI-to-ultrasound image synthesis results from different models. (a) Original axial MR images of the uterus. (b-f) Corresponding synthetic ultrasound images generated by five comparative models, respectively.
Figure 2: Visualization of selected feature maps from the modality-agnostic feature extractor (MAFE).
Metric
CycleGAN [ 44 ]
UNIT [ 24 ]
MUNIT [ 20 ]
DCLGAN [ 17 ]
SG-CycleGAN
Accuracy
0.7431 ± 0.0322
0.7465 ± 0.0049
0.5833 ± 0.0279
0.7240 ± 0.0533
0.7691 ± 0.0302
Sensitivity
0.7188 ± 0.0450
0.7882 ± 0.0547
0.3958 ± 0.1179
0.6424 ± 0.1702
0.8160 ± 0.0214
Specificity
0.7674 ± 0.0260
0.7049 ± 0.0491
0.7708 ± 0.0696
0.8056 ± 0.0650
0.7222 ± 0.0772
Precision
0.7551 ± 0.0300
0.7293 ± 0.0184
0.6340 ± 0.0272
0.7707 ± 0.0131
0.7500 ± 0.0464
F1-score
0.7362 ± 0.0364
0.7560 ± 0.0149
0.4777 ± 0.0843
0.6870 ± 0.0985
0.7802 ± 0.0192
ROC AUC
0.8177 ± 0.0178
0.8205 ± 0.0360
0.6848 ± 0.0337
0.8252 ± 0.0558
0.8416 ± 0.0188
Table 2: Comparison of Downstream Classification Task Performance (MobileNet-V2 on Synthetic Data).
Figure 3: Receiver operating characteristic (ROC) curves of the downstream invasion classifier trained on synthetic data. The lightweight MobileNet-V2 classifier was trained using synthetic ultrasound images generated by (a) CycleGAN, (b) UNIT, (c) MUNIT, (d) DCLGAN, and (e) the proposed SG-CycleGAN.
Metric
MobileNet-V2 [ 30 ]
EfficientNet [ 34 ]
MobileViT [ 26 ]
LSNet (Ours)
FLOPs
0.338G
0.513G
0.562G
0.289 G
Params
1.013M
204.5K
294.5K
391.8K
Accuracy
0.9499 ± 0.0092
0.9345 ± 0.0263
0.9688 ± 0.0080
0.9836 ± 0.0024
Sensitivity
0.9573 ± 0.0148
0.9028 ± 0.0552
0.9762 ± 0.0106
0.9950 ± 0.0014
Specificity
0.9425 ± 0.0140
0.9663 ± 0.0028
0.9613 ± 0.0084
0.9722 ± 0.0051
Precision
0.9435 ± 0.0127
0.9640 ± 0.0012
0.9619 ± 0.0081
0.9729 ± 0.0048
Table 3: Performance Comparison of the Proposed Lightweight Screening Network (LSNet) versus Benchmark Models on Synthetic and Real Ultrasound Data.
Figure 4: Comparison of receiver operating characteristic (ROC) curves among lightweight models for myometrial invasion classification.
Model Variant
Accuracy
Sensitivity
Specificity
Precision
F1-score
ROC AUC
LSNet w/o Syn. Pre-train
0.9550 ± 0.0106
0.9467 ± 0.0246
0.9633 ± 0.0143
0.9631 ± 0.0137
0.9545 ± 0.0111
0.9914 ± 0.0017
LSNet w/o Grad. Distill.
0.9617 ± 0.0042
0.9700 ± 0.0071
0.9533 ± 0.0062
0.9541 ± 0.0058
0.9620 ± 0.0043
0.9909 ± 0.0008
LSNet (Full Model)
0.9836 ± 0.0024
0.9950 ± 0.0014
0.9722 ± 0.0051
0.9729 ± 0.0048
0.9838 ± 0.0024
0.9873 ± 0.0012
Table 4: Ablation Study of the Proposed Lightweight Screening Network (LSNet).
Model Variant
LSNet w/ Full Attention ( k=N )
LSNet w/ Sparse Attention ( k=⌊N/10⌋ )
LSNet w/ Sparse Attention ( k=⌊N/25⌋ )
LSNet w/ Sparse Attention ( k=⌊N/50⌋ )
FLOPs
0.305G
0.290G
0.289G
0.288G
CPU Inference Speed (seconds/sample)
0.438 s
0.172 s
0.157 s
0.154 s
GPU Memory Usage (batch size=4)
10.1G
7.9G
7.8G
7.8G
Accuracy
0.9821 ± 0.0024
0.9812 ± 0.0031
0.9836 ± 0.0024
0.9831 ± 0.0014
Sensitivity
0.9963 ± 0.0030
0.9940 ± 0.0049
0.9950 ± 0.0014
0.9970 ± 0.0024
Specificity
0.9643 ± 0.0049
0.9683 ± 0.0014
0.9722 ± 0.0051
0.9692 ± 0.0014
Table 5: Ablation Study on Gradient-Guided Sparse Attention with Progressive Sparsity Levels.
Experiment
Accuracy
Sensitivity
Specificity
Precision
F1-score
ROC AUC
Exp #1
0.9866 ± 0.0044 (0.9777-0.9940)
0.9940 ± 0.0042 (0.9847-1.0000)
0.9792 ± 0.0078 (0.9625-0.9938)
0.9795 ± 0.0077 (0.9632-0.9939)
0.9867 ± 0.0044 (0.9772-0.9943)
0.9872 ± 0.0056 (0.9751-0.9969)
Exp #2
0.9806 ± 0.0053 (0.9702-0.9896)
0.9941 ± 0.0042 (0.9848-1.0000)
0.9671 ± 0.0096 (0.9474-0.9848)
0.9680 ± 0.0094 (0.9484-0.9853)
0.9808 ± 0.0052 (0.9699-0.9902)
0.9859 ± 0.0056 (0.9741-0.9956)
Exp #3
0.9836 ± 0.0049 (0.9732-0.9926)
0.9970 ± 0.0030 (0.9906-1.0000)
0.9702 ± 0.0093 (0.9504-0.9874)
0.9710 ± 0.0090 (0.9518-0.9878)
0.9838 ± 0.0049 (0.9736-0.9926)
0.9888 ± 0.0047 (0.9789-0.9969)
Overall Mean ± SD
0.9836 ± 0.0024
0.9950 ± 0.0014
0.9722 ± 0.0051
0.9729 ± 0.0048
0.9838 ± 0.0024
0.9873 ± 0.0012
Table 6: Bootstrap Validation Results for LSNet Across Three Independent Experiments (n=10,000 resamples per experiment)
Figure 5: Boxplots comparing Bootstrap distributions of performance metrics across three independent experiments. Each box shows the median (center line), interquartile range (box bounds), and range (whiskers) for Accuracy, Sensitivity, Specificity, Precision, F1-score, and ROC AUC.
Target Population & Prevalence (P)
PPV
NPV
LR+
Expected Screen-Positive Rate
NNS to Detect One True Positive
General Screening (P = 0.02%) [ 19 ]
0.71%
≈ 99.99%
35.79
2.80%
≈ 5025
Abnormal Postmenopausal Bleeding (P = 9%) [ 10 ]
78.00%
99.95%
35.79
11.48%
≈ 11
Atypical Hyperplasia (P = 30.8%) [ 31 ]
94.10%
99.77%
35.79
32.57%
≈ 3
Lynch Syndrome (P = 40%) [ 5 , 29 ]
96.00%
99.66%
35.79
41.47%
≈ 3
Table 7: Theoretical screening indicators of the LSNet model (Sensitivity=0.9950, Specificity=0.9722) across populations with different disease prevalences (P).
Sonographer #
Sensitivity (95% CI)
Specificity (95% CI)
ROC AUC (95% CI)
Junior group
#1
0.69 (0.65-0.73)
0.83 (0.82-0.84)
0.759 (0.749-0.770)
#2
0.79 (0.76-0.83)
0.60 (0.59-0.61)
0.698 (0.687-0.709)
#3
0.84 (0.80-0.87)
0.81 (0.80-0.82)
0.822 (0.812-0.831)
#4
0.65 (0.60-0.69)
0.80 (0.79-0.81)
0.725 (0.714-0.736)
#5
0.59 (0.55-0.64)
0.76 (0.75-0.77)
0.673 (0.662-0.685)
Table 8: Sonographer Performance in Myometrial Invasion Classification.
Figure 6: (a) CycleGAN Structure; (b) Structure-Guided Cycle-Consistent Adversarial Network (SG-CycleGAN) with Modality-Agnostic Feature Extraction and Gradient reversal
Figure 7: Gradient distillation with simulated gradient learning, guiding the student network’s sparse attention mechanism.
Figure 8: Histograms of bootstrap distributions for Experiment 1. The figure displays the frequency distributions of six performance metrics (Accuracy, Sensitivity, Specificity, Precision, F1-score, and ROC AUC) across 10,000 bootstrap resamples. The approximately normal shapes and minimal skewness of the distributions confirm the statistical reliability of the model’s performance in the first independent experiment.
Figure 9: Histograms of bootstrap distributions for Experiment 2. The symmetrical, near-normal shapes of the histograms further validate the stability and reproducibility of the model’s results in the second independent trial.
Figure 10: Histograms of bootstrap distributions for Experiment 3. The consistent, well-formed distributions with negligible skew provide additional evidence for the robustness of the proposed method.
Intelligent Medical Systems E130, German Cancer Research Center (DKFZ), Heidelberg, Germany · Department of Applied Tumor Biology, Universitätsklinikum Heidelberg, Heidelberg, Germany · Cooperation Unit F210, German Cancer Research Center (DKFZ), Heidelberg, Germany +9
Center for AI and Data Science (CAIDAS), Julius-Maximilians-Universität Würzburg, Würzburg, Germany · Institute for Computational Imaging and AI in Medicine (CompAI), Technical University of Munich, Munich, Germany · Pattern Recognition Lab, Friedrich-Alexander Universität Erlangen-Nürnberg, Erlangen, Germany +1