Uterine sarcomas are rare, occurring in roughly 5 in 10,000 women, and they carry a grim prognosis: the 5-year overall survival rate does not typically reach 50%, even when accounting for variation by histopathological subtype. The category excludes carcinosarcoma (which is epithelial in origin and carries a comparatively better prognosis), and encompasses leiomyosarcoma (LMS), which accounts for about 60% of cases, followed by endometrial stromal sarcoma (ESS), adenosarcoma, undifferentiated sarcoma, spindle-cell sarcoma, and smooth-muscle tumors of uncertain malignant potential (STUMPs). Collectively, uterine sarcomas represent only 2-3% of all uterine tumors, with the vast majority being benign uterine leiomyomas.
Why preoperative diagnosis is especially high-stakes: For benign leiomyomas, surgical options include myomectomy or total hysterectomy depending on a patient's desire to preserve fertility. For sarcomas, total hysterectomy without fertility preservation is required because morcellation can scatter tumor cells and worsen prognosis. Critically, unlike cervical or endometrial cancers, uterine sarcomas are difficult to biopsy preoperatively, making imaging the primary diagnostic tool. Misdiagnosis in either direction carries serious consequences: a missed sarcoma leads to inadequate surgery and potential dissemination, while an overdiagnosed sarcoma results in unnecessary radical surgery and loss of fertility.
Where MRI falls short: MRI is considered the most reliable preoperative imaging modality for uterine tumors due to its high contrast resolution and absence of radiation. Known sarcoma features on MRI include irregular margins and high T2-weighted imaging (T2WI) signal relative to normal myometrium, whereas leiomyomas typically show well-defined margins and low T2WI signal. However, degenerative and cellular leiomyoma variants frequently mimic sarcomas on MRI, producing overlapping imaging features that routinely lead to misclassification. Studies confirm that occult uterine sarcomas are missed as benign tumors prior to surgery with meaningful frequency.
This 2022 study from three Japanese institutions (University of Tokyo Hospital, Tokyo Metropolitan Cancer and Infectious Diseases Center Komagome Hospital, and Showa General Hospital) is the first to apply deep neural network (DNN) models specifically to the MRI-based diagnosis of uterine sarcomas. The authors frame their work as a universal framework for applying deep learning to the diagnostic imaging of rare tumors, where the scarcity of cases has historically prevented adoption of high-parameter AI architectures.
The study enrolled patients treated between 2008 and 2020 at three institutions. The final cohort comprised 63 uterine sarcoma cases and 200 uterine leiomyoma cases, totaling 263 patients. The sarcoma group included LMS (36 of 63 cases, 57%), ESS, adenosarcoma, undifferentiated sarcoma, spindle-cell sarcoma, and STUMPs. Carcinosarcomas were deliberately excluded because their epithelial origin and comparatively better response to treatment place them in a distinct clinical category. All tumors were pathologically confirmed by well-trained pathologists after surgical resection.
Inclusion and exclusion criteria: For the sarcoma group, multiple MRI scans obtained at different time points for the same patient were included to increase the volume of training images, with no restriction on the preoperative acquisition window. For the leiomyoma group, only images acquired within one year prior to resection were used. Patients were excluded if they had undergone pseudo-menopause therapy within three years before resection, had co-existing abdominal tumors, or had ovarian cysts of 3 cm or more. This exclusion strategy aimed to limit confounding features that could introduce systematic bias into model training. There were no restrictions on tumor number or diameter in either group.
MRI sequences: Fifteen MRI sequences were used, including axial T2-weighted imaging (T2axi), sagittal T2-weighted imaging (T2sag), coronal T2-weighted imaging (T2cor), diffusion-weighted imaging (DWI), apparent diffusion coefficient (ADC) maps, dynamic contrast-enhanced sequences in multiple planes, fat-suppressed T1-weighted sequences, and fat-suppressed T2-weighted sequences. DICOM data from each institution were standardized by converting to normalized JPEG format using Horos software. Six radiologists, including three board-certified specialists (14, 13, and 8 years of experience) and three non-certified practitioners (4, 4, and 2 years of experience), independently scored tumor margins and degeneration on the leiomyoma set. Degeneration or irregular margins were defined as present when at least four of the six radiologists agreed.
A notable aspect of the study design is its multi-institutional structure with heterogeneous imaging conditions: different scanners and protocols across the three sites produced imbalanced sequence availability, with some sequences present in nearly all cases (T2sag in 98.4%, T2axi in 98.0%, DWI in 96.5%) and others in far fewer patients. The authors developed a scoring system to handle this imbalance, assigning sequence-level scores of +1 for sarcoma and -1 for leiomyoma and summing across available sequences to produce a patient-level prediction.
The deep neural network selected for this study was MobileNet-V2, a relatively compact architecture consisting of 88 layers with a fixed input size of 224 x 224 pixels and 3,538,984 learnable parameters. MobileNet-V2 uses inverted residual blocks with linear bottlenecks, making it computationally efficient while retaining strong feature extraction capability. The choice of a smaller architecture was deliberate: given the scarcity of sarcoma cases, a larger network would require far more training data to avoid overfitting. The optimizer was AMSGrad, a variant of the Adam optimizer with convergence guarantees in non-stationary settings, with a learning rate of 0.0001.
Data augmentation: To address the limited dataset size (particularly the 63 sarcoma cases), aggressive data augmentation was applied. Each patient's MRI slices were augmented to produce 437,500 slices per class (sarcoma group and leiomyoma group), totaling 875,000 augmented slices. Augmentation operations included flips, rotations, zooms, and brightness adjustments. In each training epoch, 35,000 slices were randomly selected from the full augmented pool, and the model was trained for 50 epochs, processing 1,750,000 slice-exposures per trained model. This approach effectively transforms a small real-world dataset into a large training corpus, though augmented samples are not independent observations.
Cross-validation and ensemble construction: Patients were randomly divided into six groups for six-fold cross-validation, keeping slice counts as balanced as possible across folds (without balancing MRI sequence types). Using five folds for training and one for evaluation, six base models were trained per learning set. The process was repeated 24 times (M1-M24) to account for stochastic variability in DNN training, generating 144 total models (6 datasets x 24 repetitions). For evaluation, 24 ensemble sets (Ens1-Ens24) were constructed, each combining 23 of the 24 models, producing ensemble predictions that pooled the output of 138 models per ensemble set. This large ensemble approach substantially reduces prediction variance compared to any single model.
Key metric choice: Conventional accuracy is problematic for imbalanced datasets because it is dominated by the majority class (leiomyomas, n=200) and can mask poor performance on the minority class (sarcomas, n=63). The authors instead prioritized SS-Avg, defined as the arithmetic mean of sensitivity and specificity. SS-Avg penalizes models that achieve high specificity by sacrificing sensitivity, ensuring the metric reflects clinical utility in detecting both true sarcomas and true leiomyomas.
The authors evaluated all 15 MRI sequences individually for both single-model predictions and ensemble predictions. Across all 15 sequences, ensemble predictions consistently outperformed single-model predictions in SS-Avg, and the ensemble results were also more stable (lower variance across cross-validation folds). This stability advantage is clinically relevant because it suggests the final model output would be more reproducible when deployed in new clinical settings.
Top-performing individual sequences: For ensemble predictions, the three highest-ranking sequences by SS-Avg were axial T2-weighted imaging (T2axi) at 89.8%, sagittal T2-weighted imaging (T2sag) at 86.9%, and diffusion-weighted imaging (DWI) at 86.5%. For single-model predictions, the rankings were similar: T2axi (86.6%), coronal T2-weighted imaging (T2cor) (84.9%), and DWI (84.1%). The superiority of T2axi and DWI is consistent with established clinical MRI practice for uterine tumor assessment, where T2 signal intensity and diffusion restriction are among the most diagnostically discriminating features.
Why DWI matters: DWI captures the Brownian motion of water molecules within tissue, with restricted diffusion (appearing as high signal on DWI) characteristic of highly cellular or densely packed tissues such as sarcomas. Uterine leiomyosarcomas typically show restricted diffusion compared to leiomyomas due to their higher cellularity and disrupted tissue architecture. The finding that DWI ranks among the top three individual sequences confirms that the DNN model is capturing biologically meaningful signal rather than arbitrary image features.
The performance gap between ensemble and single-model predictions was consistent across all 15 sequences, not just the top performers. This generalizability of the ensembling benefit suggests it reflects genuine variance reduction through model averaging rather than sequence-specific phenomena. Even for sequences with comparatively lower individual SS-Avg values, ensembling improved both performance and stability.
The authors evaluated all 32,768 possible combinations of the 15 MRI sequences to identify the top 10 by SS-Avg. These top 10 combination sets achieved a pooled average of: accuracy 90.3%, SS-Avg 90.8%, sensitivity 89.8%, and specificity 91.7%. Combination set 1, consisting of T2axi, T2sag, and DWI, achieved the highest SS-Avg at 91.3% and accuracy at 89.9%, demonstrating that the three best individual sequences together form the most diagnostically powerful combination. The median AUC for ensemble predictions in combination set 1 was 0.9383, compared to 0.9284 for single-model predictions, confirming that ensembling provides measurable improvement even at the combination level.
Radiologist performance without AI: In the first diagnostic examination (no AI support), the three board-certified radiological specialists achieved a group average accuracy of 88.3%, SS-Avg of 82.4%, sensitivity of 71.0%, and specificity of 93.8%. The three non-certified practitioners averaged accuracy 80.1%, SS-Avg 69.6%, sensitivity 47.6%, and specificity 91.5%. The DNN models significantly outperformed both groups in SS-Avg and sensitivity (p less than 0.05 for both comparisons with specialists). Notably, sensitivity was the sharpest differentiator: the DNN achieved 89.8% sensitivity versus 71.0% for specialists and only 47.6% for practitioners, meaning radiologists, particularly less experienced ones, were missing roughly half of all sarcoma cases.
AI-assisted radiologist performance: After a one-month interval, all six radiologists repeated the diagnostic examination while viewing the DNN's sarcoma likelihood score and per-sequence results for each patient. The improvement was most striking for practitioners: accuracy rose from 80.1% to 92.3%, SS-Avg from 69.6% to 90.8%, sensitivity from 47.6% to 87.8%, and specificity from 91.5% to 93.7%. In the AI-assisted examination, practitioners actually outperformed specialists on all four metrics, though the differences were not statistically significant (p greater than 0.05). All six individual radiologists showed increases in both sensitivity and SS-Avg between the two examinations.
A behavioral observation from the study is worth noting: after AI support, radiologists who disagreed with the DNN's output spent more time reviewing the case and applied greater care to their interpretation. This suggests that the DNN serves not just as a classification tool but as a calibration signal that prompts deeper review when human and AI assessments diverge.
Detailed analysis of misclassified cases reveals important patterns in both the DNN's failure modes and its complementary relationship to human radiologist errors. For the combined MRI sequence predictions (average across top 10 combination sets), the DNN produced 6 false negative sarcoma cases (sarcoma likelihood below 50%) and 19 false positive leiomyoma cases (sarcoma likelihood above 50%). The DNN's error profile, lower false negatives and higher false positives relative to radiologists, is clinically preferable for a screening or triage tool, where missing a sarcoma carries far greater consequence than unnecessarily escalating workup of a leiomyoma.
False negatives (missed sarcomas): The 6 DNN false negatives included 3 cases of LMS, 2 cases of cervical adenosarcoma, and 1 case of STUMP. The cervical adenosarcoma failures are partially explained by anatomical differences: cervical tumors have a different spatial context than the fundal or body sarcomas that dominated the training set. One of the missed cervical adenosarcomas was only approximately 3 cm in diameter, providing limited MRI signal for the model. STUMPs are pathologically intermediate lesions that are inherently difficult to diagnose even by expert pathologists, so the DNN's failure on this case reflects the fundamental ambiguity of the entity rather than a model deficiency.
False positives (overcalled leiomyomas): Among the 19 false positive leiomyoma cases, 12 (63.2%) showed degeneration, and only 1 (5.3%) had irregular margins. Tumor diameter showed no systematic bias in the false positive group. This pattern suggests that the DNN is learning sarcoma-like internal heterogeneity, including degeneration-associated signal changes, as a positive predictor, which is biologically sensible because sarcomas also show internal heterogeneity. Degenerative leiomyomas, which mimic this appearance, therefore confuse the model in the same way they confuse human radiologists.
The complementary error patterns of the DNN and radiologists have practical implications for a combined human-AI workflow. When the DNN's errors are compared to radiologist errors case by case, the two sources of error do not perfectly overlap. Some cases misdiagnosed by the DNN were correctly identified by radiologists, and vice versa. A sequential review protocol in which radiologists focus highest attention on cases where their initial reading diverges from the DNN output could capture benefits from both sources.
Small dataset from a rare disease: With only 63 sarcoma cases available across three institutions over a 12-year period, the training set is inherently limited for a high-parameter model. The authors address this through augmentation, ensemble averaging, and the SS-Avg metric, but none of these strategies can fully substitute for a larger real-world case pool. The model was evaluated only through cross-validation on the same three institutions used for training, meaning there is no independent external validation set. Cross-validation performance consistently overestimates generalization accuracy, and the true performance on prospectively collected cases from a new institution remains untested.
Imbalanced sequence availability: The multi-institutional design introduced an inherent imbalance: some sequences had near-complete availability (T2sag: 98.4%, T2axi: 98.0%, DWI: 96.5%) while others were present in far fewer patients. The top three performing sequences also happen to be those with the greatest coverage. The authors acknowledge that this availability bias may inflate the apparent performance of T2axi, T2sag, and DWI compared to what would be observed in a dataset where all sequences were acquired for all patients.
Binary output without histopathological subtyping: The DNN was designed to classify images as either uterine sarcoma or uterine leiomyoma. It does not distinguish among LMS, ESS, adenosarcoma, undifferentiated sarcoma, or STUMP, all of which have different surgical and systemic treatment implications. A model that could predict histopathological subtype from MRI alone would be clinically far more valuable, but the rarity of individual subtypes makes subtype-specific training currently infeasible.
Absence of other abdominal tumors: The exclusion of patients with co-existing ovarian tumors and ovarian cysts above 3 cm means the model was trained on a relatively clean anatomical context. In clinical practice, uterine leiomyomas frequently co-exist with ovarian endometriomas and other pelvic pathology that could introduce confounding features on MRI. Extending the model to handle these co-existing lesions would require a substantially more complex study design. Additionally, the absence of clinical covariates (age, serum LDH, tumor markers, CA-125) means the model does not capture information that radiologists routinely integrate into their interpretations.
Scaling the rare-disease approach: The authors position MobileNet-V2 with augmentation and ensembling as a transferable framework for any rare tumor where institutional datasets are too small for conventional deep learning. The core strategy, training a compact architecture with extreme augmentation and pooling predictions across many independently trained models, can be applied to other rare gynecologic or soft-tissue tumors where curated large datasets do not exist. Demonstrating that acceptable diagnostic accuracy is achievable with 63 positive cases is itself a methodological contribution to the field.
Multiparametric expansion with clinical variables: The discussion explicitly identifies inclusion of additional clinical information, specifically age, laboratory data, and tumor markers, as a necessary next step for achieving the diagnostic granularity needed for histopathological subtype prediction. A multiparametric model combining MRI features extracted by the DNN with clinical and laboratory variables could approach the performance of molecular pathology tools for sarcoma subtyping, potentially reducing unnecessary surgical exploration or enabling more targeted biopsy guidance.
Interpretability and Grad-CAM analysis: The authors note that visualizing how DNN models make decisions is critical for identifying areas for improvement. Gradient-weighted class activation mapping (Grad-CAM) and similar saliency methods can generate heat maps highlighting which regions of an MRI slice most strongly influenced the prediction. If the model's attention maps consistently highlight clinically recognizable features (irregular margins, high T2 signal, restricted diffusion zones), this would increase clinician trust and facilitate validation against known MRI criteria. The identification that false positives cluster around degenerative change suggests that Grad-CAM maps for these cases would localize to internal heterogeneity zones.
Prospective clinical integration: The current study uses retrospective data with a one-month washout between diagnostic examinations designed to minimize recall bias, but this design cannot replicate real clinical workflow. A prospective study embedding DNN output into the radiology reporting system, where radiologists see the sarcoma likelihood score as a standard part of the MRI read, would provide the most direct evidence of clinical utility. Such a trial would also allow comparison of downstream clinical decisions, including referral patterns, surgical planning, and time to definitive diagnosis, between AI-assisted and standard reads.