Uterine tumors, which include both benign and malignant tumors, create a substantial healthcare challenge because of their high occurrence rates. Uterine cancer constitutes the most frequently occurring malignancy in women throughout the United States, which results in a 3.4% lifetime cancer risk (1). The sixth most prevalent cancer in 2012 (2) affects breast cancer as the most common cancer among women in the United States. The most common tumor type that affects women is leiomyoma (3). The tumors develop during the female reproductive life along with the perimenopausal period, thus creating debilitating symptoms of hemorrhage, pelvic heaviness, and pain (4). The malignant varieties of uterine cancers include endometrial cancer along with uterine sarcomas, which have a poor prognosis despite being a rare type of cancer.

Uterine cancers develop among women who experience advanced age and reach menopause later and who show hormonal imbalances and who have never given birth (5). The relationship between genetic pre-dispositions such as Lynch syndrome and obesity, which causes increased estrogen levels, which can lead to tumor development, serves as an additional risk factor (6). Researching the risk factors associated with uterine cancer proves necessary because of rising cancer rates and growing death rates which reached 1.1 increase from 1999 to 2016 with black women experiencing higher rates than white women (7). The medical establishment achieves its mission to decrease patient death rates while delivering better patient care through the practice of diagnosing diseases at their earliest stages and providing prompt medical intervention.

Deep learning algorithms have been applied for the evaluation and interpretation of images obtained from CT scans with the objective of improving diagnostic accuracy. For instance, they have been applied to help in distinguishing between benign and malignant uterine smooth muscle tumors (8). Deep learning enables automatic tumor segmentation which has become a vital component of treatment planning. The ability of deep learning in improving the diagnostic accuracy of CT images has been proven in studies involving ovarian cancer (9), which further supports its application for the evaluation of uterine smooth muscle tumors. The deep learning method for staging endometrial cancer using MRI images has produced positive results according to the research findings (10). The applications face multiple difficulties because of dataset discrepancies which create the main obstacles. Future studies should develop models which exhibit reduced variations while including multimodal imaging elements according to (11).

Although there are significant advances in the employment of deep learning techniques in medical image analysis, there are still challenges to be encountered in diagnosing uterine cancer from computed tomography images. Most previous works have harnessed CNN architectures that are mostly designed and trained for extracting local features, with limited representation of the long-range contextual relations within the image. Thus far, only a limited number of studies have focused on the classification of uterine cancer CT images into normal, benign, and malignant categories with high accuracy, using realistic datasets from clinical settings. In general, the following contributions can be summarized as part of the research study in uterine cancer diagnosis using artificial intelligence (AI):

The subsequent sections are arranged as follows: a thorough analysis of the pertinent research on uterine diseases is given in Section 2. In Section 3, the dataset is described, the applied methodology is described, and the structure of the suggested model is explained in depth. The experimental outcomes of the suggested model are shown in Section 4. The scalability and comprehensiveness of the suggested model are covered in Section 5, and the results are summarized and future research opportunities are outlined in Section 6.

Deep learning models used for many types of medical imaging, including computed tomography, magnetic resonance imaging, and histology imaging, have progressively helped diagnose uterine cancer and classify benign and malignant tumors. The most recent findings from studies on using AI for uterine cancer classification are presented in this section.

The study presented in (12) introduced a modified YOLOv5 framework for the classification of uterine abnormalities based on ultrasound images. The proposed study focused on the redesigning of the head module of the original YOLOv5 architecture, the addition of global attention mechanism (GAM), and the use of a modified activation function and ResNeXt-based CSP blocks to create a new backbone to capture the very subtle features from the noisy ultrasound images more effectively. A dataset of 3,026 clinically-acquired ultrasound images was used for model training and validation, and the images were labeled into the classes: cervical cyst, uterine fibroid, and normal. Experimental results showed that the proposed method achieved an accuracy of 80.5%, indicating the effectiveness of architectural enhancements and attention processes in enhancing the classification of uterine abnormalities based on ultrasound.

In another research work (13), the goal is to offer an objective evaluation regarding the aspects influencing the success of UFE and to promote the use of an interpretable machine learning (ML) approach to aid in decision-making in the clinical setting via features extracted from pre-operative MRI. A carefully filtered dataset with 74 patients and 311 fibroids was created, and they used Deep Set Networks, which allow for permutation-equivariant aggregation of features over the collection of fibroids. Accuracy values of 81–88% (AUC = 0.81–0.87) were measured in individual symptoms for the proposed models, as well as 75% (AUC = 0.74) for the overall clinical success, whereas at the fibroid level, the prediction task by ensemble tree-based techniques resulted in an accuracy of 76% max.

Another study (14) presented the role of AI in the diagnosis and treatment of uterine fibroids and uterine sarcomas, covering studies published between 2019 and March 2025. The research examined various AI techniques which included radiomics machine learning and deep neural network systems that worked with ultrasound and MRI medical images. The methods successfully distinguished between benign leiomyomas and malignant leiomyosarcomas which helped doctors develop treatment plans and improved results from minimally invasive procedures including HIFU and uterine artery embolization. AI systems demonstrated better performance than expert radiologists in most cases according to the results while multiple studies reached diagnostic accuracy levels above 0.85 AUC scores.

In the study (15), a framework was proposed for surgical decision-making in uterine fibroid management through integrating female sex hormone levels with fibroid characteristics. The study included 618 women diagnosed with uterine fibroids (UFs) from a multicenter hospital, of whom 238 underwent surgery. Multiple supervised ML algorithms, such as support vector machine (SVM), decision tree (DT), random forest (RF), logistic regression (LR), and k-nearest neighbors (KNN), with 126 different input combinations derived from hormonal markers (FSH, LH, E2, PRL, and AMH) and morphological variables. The RF model achieved the highest accuracy of 91% and an AUC of 0.88 using LH, FSH, E2, and AMH. The model showed a high level of clinical concordance through external validation using 20 independent cases, achieving a 90% agreement rate with a blinded gynecologist.

As per the study (16), a multicenter deep learning-based framework for automatically outlining clinical target volumes (CTV) and planning target volumes (PTV) in radiotherapy for uterine cancers. The research employed the self-configuring nnU-Net architecture to solve manual contouring challenges which included observer differences and institutional result variations and heavy workload demands. The researchers collected a dataset which contained 602 contrast-enhanced CT scans that included cervical and endometrial cancer cases from multiple medical centers. The researchers conducted a detailed evaluation of three nnU-Net configurations which included 2D slice-level and 3D full-resolution and 3D cascaded models through testing with Dice Similarity Coefficient (DSC) and Average Surface Distance (ASD) and Hausdorff Distance (HD95). The 3D full-resolution nnU-Net provided the best overall results, with average DSC values of 83.42% for PTV and 81.23% for CTV in internal testing. Additionally, evaluations from experienced radiation oncologists showed that approximately 90% of the automatically created contours required no changes or only minor adjustments, demonstrating the clinical usefulness of the proposed method. While the study (17) introduced deep learning-based nnU-Net models, for automatic segmentation of uterine fibroids and their surrounding organs based on MRI images to support high-intensity focused ultrasound HIFU surgery planning. Used a retrospective dataset of 550 T2-weighted MR images. The evaluation showed that the proposed 3D nnU-Net significantly outperformed state-of-the-art methods such as HIFUNet, U-Net, R2U-Net, ConvUNeXt, and 2D nnU-Net, achieved DSC of 92.55% for the uterus, 89.63% for the endometrium, 90.45% for the urethral orifice, 97.75% for the bladder, 95.63% for fibroids, and 92.69% for the spine.

In a study (18) a novel deep learning-based 3D super-resolution DWI (SR-DWI) radiomics model was proposed for predicting the prognosis of high-intensity focused ultrasound (HIFU) ablation of uterine fibroids. Radiomics features were extracted from manually segmented fibroid regions and subsequently reduced using t-test, Pearson's correlation, and the LASSO regression algorithm. machine learning classifiers, including SVM, RF, and LightGBM, were trained and validated on multicenter datasets with both internal and external testing cohorts. All DWI radiomics models showed superior AUC, according to experimental results. The best-performing HR-DWI model (SVM) achieved an AUC of approximately 0.805 in internal testing, whereas the SR-DWI–based models demonstrated better performance, with AUC values of 0.876 in internal testing and 0.800 in external validation, indicating a statistically significant improvement over HR-DWI (P < 0.05). The findings show the potential of combining deep learning-based super-resolution imaging with radiomics. This approach can improve pre-operative prognostic assessment of eligible candidates for HIFU therapy.

According to the study (19) an image processing-based solution was proposed for the diagnosis of cervical cancer from uterine cervix images using transfer learning architectures to reduce workload and assist experts by leveraging deep convolutional neural networks. Histogram equalization was applied to enhance image contrast before classification, while to suppress noise, Gaussian filtering was used. Many transfer learning models, including AlexNet, MobileNetV2, DenseNet201, ResNet50, VGG19, and Xception, were systematically evaluated using a 10-fold cross-validation strategy on the Herlev pap-smear dataset. The experimental results showed that the VGG19 model achieved the best performance, with an accuracy of 98.26%, outperforming other models on the same dataset.

Another study (20) proposed a machine-learning-based radiomics framework using pre-operative contrast-enhanced computed tomography (CECT) images to differentiate uterine leiomyomas from leiomyosarcomas. After standardized pre-processing, PyRadiomics was used to extract the radiomic features, and three methods (Boruta, LASSO, and RFE) were used to select features, with multiple classifiers such as GLM, RF, and SVM. The diagnostic performance of the proposed models reached strong results because they achieved test data AUC results between 0.78 and 0.97 which exceeded the performance of radiologists who achieved AUC results between 0.70 and 0.78 when the patient's clinical data were available. The study demonstrates how CT-based radiomics can function as a decision-support system during pre-operative evaluations. The EfficientNetB0 model achieved classification with an accuracy of 99% in the study (21) which used 1,990 ultrasound images to automatically classify uterine fibroids into two categories. The study employed an attention-based model which used attention mechanisms to guide the model towards important clinical areas while it neglected non-essential parts of the body. Table 1 presents a comparison of previous research on uterine diseases.

Hybrid deep learning systems development has demonstrated potential to improve medical imaging techniques and cancer detection methods. Alswilem and Pacal (22) conducted a comprehensive comparative study which assessed both computational efficiency and diagnostic accuracy of deep learning models used for automated breast cancer detection in ultrasound imaging. The study found that RexNet-200 achieved optimal performance with minimal computational resources which included 13.81 million parameters and 3.05 GFLOPs while maintaining 95 percent accuracy. The study demonstrates how model complexity needs to achieve a specific level which medical professionals need for their work throughout our DenseNet121-Transformer. Demirtaş Alpsalaz et al. (23) colleagues created a hybrid model which combines EfficientNet-B3 with Vision Transformer to detect colon cancer through their attention fusion mechanism and the system reached 96.2 percent accuracy while achieving an MCC score of 0.961. Their work demonstrated that attention-based fusion effectively harmonizes local texture features extracted by CNNs with global contextual dependencies captured by transformers, directly supporting our implementation of multi-head self-attention for feature integration. Çakmak (24) examined various machine learning techniques to improve the diagnostic process of hematological disorders and achieved 98.38 percent accuracy with LightGBM while showing that selecting features and optimizing models represent crucial components for building clinical decision support systems. Çakmak and Pacal (25) performed a study that compared four CNN architectures (ResNet50, DenseNet169, InceptionV3, InceptionV4) to classify breast ultrasound images and found that InceptionV3 achieved the best results with 96.67% accuracy and 96.55% precision. Their findings demonstrate that densely connected architectures together with multi-scale feature extraction systems work effectively which serves as our basis for selecting DenseNet121 to function as our convolutional current architecture. Alpsalaz et al. (26) created a deep learning model based on MaxViT which they used to classify Alzheimer's disease through MRI scans and achieved 99.60% accuracy by implementing transfer learning together with multi-axis attention mechanisms. This research shows how vision transformers successfully capture medical image details through their ability to detect both small and large image features which supports our choice to use transformer-based self-attention in our uterine cancer classification system.

Recent studies show that deep learning techniques successfully analyze medical images across different clinical settings. The medical field uses convolutional neural network approaches to detect and classify tumors which result in accurate brain tumor diagnosis through magnetic resonance imaging (27). The advanced object detection frameworks which include YOLO-based models successfully detect abdominal diseases because they use advanced feature extraction and pre-processing methods to handle complicated anatomical areas (28). The research demonstrates that data pre-processing and imbalance handling methods serve as essential elements which enhance machine learning models accuracy for medical diagnostic purposes (29). The research demonstrates that hybrid CNN-Transformer architectures with attention mechanisms work effectively in medical imaging applications because they provide strong theoretical and empirical foundations which support our proposed methodology.

This section will go over the findings of the study's suggested model, which was tested using the KAUH-UCCTD uterine tumor datasets that were gathered from King Abdullah University Hospital in Jordan. Each model was trained using the same set of parameters: 50 epochs, learning rate of 0.0001, Adam activation function, batch size 16, image size (224,224), and class cross-entropy loss function. The dataset is divided into three subsets for this study: 10% for testing, 10% for validation, and 80% for training. Additionally, a Jupyter laptop and an RTX 3050 GPU were used to train the models locally.

The section displays all model performance metrics which include accuracy, precision, sensitivity, and specificity together with their F1 score and AUC. The KAUH-UCCTD dataset evaluates proposed model performance which researchers use to compare their model against five existing models: VGG16 and VGG19 and MobileNetV2 and DenseNet121 and ResNet50. Table 3 shows that the proposed model achieved better results than all baseline models. The system proved itself effective for uterine cancer detection because it achieved all evaluation metrics which included 87.44% accuracy and 87.48% precision and 95.20% specificity and 99.41% AUC and 87.17% F1-score. The DenseNet121 architecture among baseline models achieved second place because it reached 84.80% accuracy and 92.40% specificity which showed its ability to extract and classify complex CT image features. The VGG19 and ResNet50 models performed worse than other models because they achieved accuracies of 80.17% and 75.33% respectively. The models examined in the study show their effectiveness through the results which Figure 3 displays.

The proposed model shows excellent performance because it can effectively distinguish three uterine tissue classes at all classification threshold points. The system evaluates its classification performance through a specific decision point which does not accurately represent its ability to rank results in multi-class situations with unbalanced classes. The selected threshold for classification results in misclassifications because of slight visual overlap between benign and malignant cases which leads to decreased accuracy although the system shows strong ability to differentiate between the two groups. The system shows high area under the receiver operating characteristic curve results but its actual accuracy performance remains lower than expected.

The results from the proposed model and the standard DenseNet121 model show their performance differences through their classification accuracy and loss measurements, which they produced during training and validation through 50 completed training epochs, as a shown Figure 4. The proposed model in the top row achieves full training accuracy through controlled training that achieves 90% accuracy while maintaining small differences between training and validation accuracy assessment. The results indicate that the model can generalize well because it achieved good results while maintaining control of overfitting problems. The proposed model's loss curve demonstrates a continuous training loss reduction while the validation loss maintains stable low levels which experience minor changes because medical CT data shows natural variability. The bottom row displays the results of DenseNet121 testing. The model achieves satisfactory training accuracy but shows poor performance during validation because it has a major gap between training results and validation outcomes. The model shows difficulties because it cannot capture all the worldwide connections that exist in images according to the high validation loss results which fail to decrease at a proper rate. The results demonstrate that using DenseNet121 with Transformer and self-attention in the proposed model has improved training stability and brought better training validation convergence results which increased generalization capabilities according to the better quantitative results that the proposed model achieved on the KAUH-UCCTD dataset.

The Figure 5 graphs below show the confusion matrices for all comparative models in classifying uterine CT images in KAUH-UCCTD in terms of VGG16, VGG19, MobileNetV2, DenseNet121, ResNet50, and our design model. The problem of identifying Benign vs. Malignant can be seen in traditional models such as VGG16 and VGG19; they are highly confused in both models, as reflected in high off-diameter points. While improvements are made in DenseNet121 and MobileNetV2 designs, but are mistaken in some cases regarding malignancies. The model presented does the best job in having the largest number of correct predictions on the main diameter, along with a considerable decrease in errors, especially in the Malignant class, exhibiting a high level of sensitivity in malign predictions and a decrease in false positives. Moreover, the model does a near-perfect job in predicting the Normal class with a negligible level of interference in the result, depicting a high level of discrimination between healthy and infected regions. The result-making matrices together ensure that the combination of the DenseNet121 model and the Self-Attention mechanism (Transformer model), in fact, increases the efficiency level in distinguishing the three classes effectively and makes the model more reliable in clinical practices.

False-negative predictions in the malignant class represent a particularly critical limitation in cancer diagnosis, as they may lead to delayed clinical intervention and adversely affect patient outcomes. Although the proposed model demonstrates high overall sensitivity and specificity, a small number of malignant cases were misclassified, primarily due to morphological similarities between benign and malignant uterine tissues. In a clinical decision-support context, such a system should be used as an assistive tool rather than a standalone diagnostic solution, with final decisions made by experienced radiologists. Moreover, future work will focus on reducing false-negative rates by incorporating cost-sensitive learning strategies, threshold optimization, or ensemble-based approaches that prioritize malignant case detection

This study presented an efficient hybrid deep learning architecture that was demonstrated, comprising the combination of DenseNet121 and the self-attention mechanism based on the Transformer for the classification of uterine CT images into normal, benign, and malignant images based on the KAUH-UCCTD dataset. In the study, the proposed hybrid model efficiently combined the strength of convolutional neural networks in identifying local textual features in the images along with the ability of the Transformer encoder to handle the images globally. The results obtained from the experiment have shown that there is superiority in the proposed method over different advanced methods such as VGG16, VGG19, MobileNetV2, DenseNet121, and ResNet50 based on different parameters such as accuracy, sensitivity, specificity, F1-score, and AUC.

Future work will be done by extending the study to validate the proposed framework on multi-institutional multi-center datasets, which improves the generalizability and clinical reliability of the framework. The inclusion of 3D volumetric information and interslice relationships from CT examinations could lead to even better diagnostic performance. The integration of XAI techniques, such as attention maps or visualization saliency, would make the model more interpretable and increase clinician trust in the models. Other future works may involve real-time deployment in hospital environments, optimization for computational efficiency, and integration with clinical workflows for practical decision-making during routine medical practice.