This study uses the publicly available Multi-Modality Ovarian Tumor Ultrasound (MMOTU) dataset (22), created by Zhao et al. at Beijing Shijitan Hospital, Capital Medical University. The MMOTU dataset has been developed to support research focused on ovarian tumor segmentation and classification across various imaging settings. It consists of ultrasound images of several devices, in both B-mode conventional ultrasound (OTU_2D) and contrast-enhanced ultrasound (OTU_CEUS) modes, spanning a wide range of texture variations and intensity distributions. The dataset contains 1,469 two-dimensional (2D) grayscale ultrasound scans in the OTU_2D subset and 170 images in the CEUS subset, each with expert radiologist annotations. The images are also accompanied by a pixel-wise binary mask outlining the tumor region and a global class label indicating whether the tumor is benign or malignant. The dataset is suitable for supervised learning, semantic segmentation, and unsupervised cross-domain adaptation tasks, given these annotations. The MMOTU dataset is rich in metadata, including the acquisition modality, image resolution, and scanner source, which facilitates model generalization and domain transfer. Each image was de-identified and is available for research use via the official GitHub repository (23). The variety of imaging modalities and diagnostic presentations in the dataset makes it a standard resource for analyzing ovarian tumors in medical imaging and computer vision. The summary of this dataset is provided in Table 1.

The Data Ingestion phase serves as the foundational stage of the proposed framework, responsible for acquiring raw ultrasound data, organizing heterogeneous inputs, validating annotations, and generating structured metadata required for reliable deep learning training. This phase ensures that all subsequent processing steps operate on a consistent, traceable, and quality-verified dataset.

The first phase of the proposed framework involved ingesting and systematically exploring the ultrasound tumor dataset to organize it, make it available, and prepare it for deep learning processing. The dataset was programmatically retrieved from cloud storage and mounted into the local processing environment, where directory paths were explicitly defined for raw ultrasound images, segmentation masks (annotations), and output logs. All intermediate and processed data were stored in organized directories to enable traceability throughout subsequent phases. An automated directory traversal and validation procedure was employed to identify and index all available data files. The system recursively searched predefined image and annotation folders, with a fallback scan of the root directory to ensure no files were omitted. This process retrieved 1,469 image files and 4,407 mask files. File path previews were generated to verify directory consistency and dataset completeness.

Label information was extracted from text files that mapped image identifiers to class labels, and the data were converted into structured metadata, enabling straightforward linking of each image to its class label. To ensure correct correspondence between ultrasound images and their segmentation masks, a robust lookup table was constructed using filename stem matching, incorporating exact, prefix-based, and containment-based matching strategies. This procedure ensured that all valid masks were correctly associated with their corresponding images. For each image, descriptive metadata including spatial resolution (width and height), number of associated masks, and pixel intensity statistics (mean and standard deviation) were computed. These attributes were consolidated into a centralized manifest file that functioned as a comprehensive dataset index and quality-control reference.

Exploratory analysis measured the mask coverage ratio, defined as the proportion of tumor pixels relative to the total image area, to identify samples that were under- or over-annotated. Image quality assessments based on pixel intensity histograms were used to detect low-contrast, overexposed, or noisy images. Additionally, patient-level identifiers embedded in filenames were analyzed to assess inter-sample dependencies. As no repeated patient instances were detected, a stratified random-splitting strategy was adopted rather than patient-wise grouping. The dataset was partitioned into training, validation, and test subsets using a 70:15:15 ratio, with stratified sampling to preserve class distributions across splits. Each subset retained its corresponding metadata records to maintain full traceability between raw inputs, annotations, and experimental outcomes. Exploratory Data Analysis (EDA) visualized class distribution Figure 2a, mask coverage Figure 2b, image size variability Figure 2c, and pixel intensity characteristics Figure 2d. Finally, qualitative validation was performed by superimposing segmentation masks onto their corresponding ultrasound images to confirm spatial alignment and annotation accuracy visually. This phase ensured the dataset is clean, consistent, and reliable for model training, providing crucial insights into class balance, mask representation, and image characteristics.

Figure 2a shows the distribution of classes and the imbalance between the classes. Specifically, classes 0 and 2 have the largest sample sizes, followed by classes 5 and 1, which have moderate representation, whereas classes 6, 3, 4, and 7 have relatively few samples. This uneven distribution indicates a class-imbalanced dataset, which could bias the model toward the majority classes during training and negatively affect its performance on underrepresented classes. In contrast, Figure 2b depicts the mask coverage distribution, which shows the proportion of each image covered by a mask. The histogram shows that most images have low mask coverage (below 0.1), while only a few exhibit higher coverage. This right-skewed distribution suggests that the regions of interest occupy a small portion of most images. Such sparsity may make it challenging for the model to effectively learn from limited labeled areas, underscoring the need for strategies such as data augmentation, loss weighting, or focused sampling to improve training balance and segmentation accuracy. Figure 2c shows the distribution of image pixel counts, indicating that the dataset contains images with varying resolutions. The multiple peaks in the histogram indicate distinct image-size groups, suggesting that the data were collected from different sources or imaging settings. This variability in image dimensions implies that resizing or normalization is necessary before model training to ensure consistent input sizes and stable learning behavior. Figure 2d depicts the relationship between image mean intensity and standard deviation, showing a positive correlation. Images with higher mean intensity generally exhibit greater pixel-value variation, indicating greater contrast or richer texture. In comparison, darker images (with lower mean intensity) tend to have lower standard deviation. This relationship provides insight into the overall brightness and contrast diversity in the dataset, which can influence model generalization and robustness.

The second phase focused on refining and standardizing the dataset produced in Phase 1 to ensure uniformity and quality before model training. The manifest file, which contains paths to images, masks, and class labels for all 1,469 samples, was used for structured preprocessing. The workflow emphasized region-of-interest (ROI) extraction, intensity normalization, and size standardization to enhance focus on tumor regions while minimizing background variability. Among three ROI modes, Mode A (crop + pad + resize), Mode B (mask background), and Mode C (dual-stream saving), Mode B was selected for this study, wherein non-tumor regions were masked out to highlight the lesion area. Each image was resized with aspect-ratio preservation and center-cropped to 224 × 224 pixels. Intensity values were clipped to the 1st-99th percentile range to enhance contrast and normalized to a fixed scale. Multiple binary masks corresponding to a single image were merged into a unified mask, ensuring complete lesion representation. Subsequently, the image was cropped around the ROI bounding box and padded to a square shape if necessary. Processed images were saved in the Phase 2 output directory using standardized naming conventions. Metadata comprising image identifiers, original and processed paths, class labels, and ROI mode were logged systematically. For Mode C, an additional full-image version was retained for dual-branch training. This phase produced a clean, intensity-normalized, and ROI-focused dataset optimized for deep learning applications.

Phase 3 expanded the dataset's diversity and addressed inherent class imbalance while preparing training-ready data pipelines. Configuration parameters such as image size, batch size, and computation device were specified, and the processed manifest and split files were loaded to construct consistent train, validation, and test sets. A multi-stage augmentation pipeline was implemented, combining geometric transformations (rotation, scaling, flipping) with photometric adjustments (brightness, contrast, gamma correction). Advanced strategies such as MixUp and CutMix were incorporated to further improve generalization and model regularization. Augmentations were applied probabilistically to training samples, while validation and test data underwent only deterministic resizing and normalization to maintain evaluation integrity. To mitigate class imbalance, class frequencies were analyzed, and corresponding weights were computed as the inverse square root of class occurrence. Weighted sampling ensured proportional representation of minority classes within each mini-batch. Additionally, weighted Cross-Entropy and Focal Loss functions were defined to reduce bias toward dominant classes. Focal Loss dynamically downweights easy samples, thereby improving the sensitivity to the minority class. Augmented image samples were visually verified for realism and correctness. A preliminary evaluation using a lightweight pretrained backbone confirmed proper data formatting, augmentation consistency, and label alignment. This phase established a robust, balanced, and model-ready dataset foundation through controlled augmentation and reweighting strategies.

Phase 4 introduced the proposed dual-branch deep learning framework, EfficientOvaNet, which leverages an EfficientNet-B3 backbone as its computational core. EfficientOvaNet does not introduce a novel backbone architecture; rather, it employs the EfficientNet-B3 model within a dual-branch design to integrate local and global contextual features. The Phase 2 dataset and the stratified partitions from Phase 3 were used for training (70%), validation (15%), and testing (15%). All inputs were standardized to 224 × 224 resolution with runtime augmentations applied to the training set. Class imbalance was handled by reweighting based on the inverse square root of frequencies, and Focal Loss was selected for its ability to emphasize complex and minority-class examples.

EfficientOvaNet consists of two synchronized convolutional branches: the first processing cropped ROI images to learn fine-grained tumor boundaries, and the second handling full ultrasound frames to capture global anatomical context. Both branches share an EfficientNet-B3 backbone with synchronized weights. Extracted feature maps from both streams undergo adaptive average pooling and concatenation before being passed through a fully connected classification head composed of normalization, activation, dropout, and an output layer representing eight tumor classes. Model optimization used an adaptive optimizer with cyclical learning rate scheduling to achieve stable convergence. Mixed-precision training accelerated computation, while dropout and early stopping provided additional regularization. The resulting model effectively fused local and contextual cues, achieving efficient dual-stream feature integration and high discriminative power.

Comprehensive validation and evaluation of EfficientOvaNet were performed to assess its stability, generalization, and clinical readiness. Table 2 summarizes the five-fold cross-validation performance of the proposed model on the MMOTU dataset. A stratified five-fold cross-validation (K = 5) scheme was employed to ensure equitable class representation. The results demonstrate stable, consistent performance across all folds, with mean accuracies and F1-score of 0.919 and an AUC of 0.982, respectively. The low standard deviations indicate minimal variance across folds, confirming the model's robustness and reliable generalization. Moreover, the narrow 95% confidence intervals further validate the model's statistical stability, underscoring its strong discriminative capability for classifying ultrasound images of ovarian tumors.

Figure 3 shows the confusion matrix, which illustrates model performance for each class. Diagonal elements indicate correctly classified samples, while off-diagonal elements show misclassifications. This figure is important because it highlights specific class confusions and the overall classification strength, providing insight into areas that may require improvement. The matrix comprises four classes labeled 0, 1, 2, and 5, which denote distinct ovarian tumor categories in the MMOTU dataset. The diagonal elements indicate correctly classified instances for each class: 306 for class 0, 207 for class 1, 298 for class 2, and 253 for class 5, reflecting high classification accuracy across all categories. The off-diagonal values represent misclassified instances, highlighting areas where the model confuses one class with another (e.g., in the actual class 0 row, 12 samples were predicted as class 1, 3 as class 2, and 15 as class 5). The color intensity in the matrix visually emphasizes prediction frequency, with darker shades indicating higher counts and lighter shades indicating fewer predictions. Overall, the predominance of higher diagonal values and minimal off-diagonal errors indicates that EfficientOvaNet achieves strong discriminative performance and reliable classification accuracy across tumor classes.

Figure 4 presents the normalized confusion matrix, which communicates relative performance across classes regardless of sample size. This visualization provides an intuitive understanding of class-wise accuracy proportions, highlighting the model's robustness and consistency across imbalanced data. This graph illustrates the proportional performance of the proposed EfficientOvaNet model across tumor classes 0, 1, 2, and 5. Each cell represents the fraction of predictions relative to the total actual instances of a class, enabling a balanced comparison regardless of class size. The diagonal values indicate class-wise accuracies of 91% for class 0, 95% for class 1, 89% for class 2, and 95% for class 5, demonstrating strong overall classification performance. Off-diagonal values highlight minor misclassifications, primarily between classes 0, 2, and 5.

Figure 5 shows ROC curves, which are crucial for understanding the trade-off between sensitivity and specificity for each tumor class. The AUC values indicate the model's discriminative power, with higher curves reflecting better performance. This figure demonstrates the reliability of EfficientOvaNet in distinguishing between classes across varying decision thresholds. This evaluates the classification performance across tumor classes 0, 1, 2, and 5. The x-axis represents the False Positive Rate (FPR), the proportion of negative instances incorrectly classified as positive. In contrast, the y-axis represents the True Positive Rate (TPR), also known as Recall or Sensitivity, the proportion of actual positives correctly identified. The dashed diagonal line from (0,0) to (1,1) represents a random classifier with no predictive ability (AUC = 0.5). Each colored curve corresponds to a specific class and shows the trade-off between TPR and FPR across different threshold values. The AUC values, as shown in the legend, are 0.97 for class 0, 0.99 for class 1, 0.98 for class 2, and 0.99 for class 5, indicating excellent discriminative ability. The curves lie well above the diagonal, indicating highly accurate, favorable rates and low false-positive rates across all classes. In particular, classes 1 and 5 achieve near-perfect classification with an AUC of 0.99.

Figure 6 presents validation accuracy and F1 score over 12 training epochs, conveying the model's learning progression and consistency between overall correctness and precision-recall trade-off. This figure is important because it demonstrates effective training dynamics and generalization capability. The x-axis represents the number of epochs, while the y-axis denotes the metric values ranging approximately from 0.67 to 0.80. Validation accuracy (blue line with circles) measures the proportion of correctly predicted instances, and the validation F1 score (orange line with squares) represents the harmonic mean of precision and recall, balancing false positives and false negatives. Both metrics begin at approximately 0.68 in the 1st epoch and steadily increase as training progresses, with minor fluctuations between the 3rd and 12th epochs. By the final epoch, both metrics approach 0.80, indicating strong model performance. The close alignment between validation accuracy and F1 score demonstrates that the model maintains a consistent balance between overall correctness and the trade-off between precision and recall, reflecting effective learning and generalization during training.

Figure 7 shows training and validation loss over 11 epochs, indicating the model's convergence behavior. The figure highlights differences between seen and unseen data, guiding readers on generalization and potential overfitting concerns. The x-axis represents the number of epochs, while the y-axis denotes the corresponding loss values ranging from 0 to 0.6. The training loss starts at approximately 0.6 and decreases rapidly during the initial epochs, gradually approaching zero by the 11th epoch, indicating effective learning and a strong fit to the training data. In contrast, the validation loss begins at approximately 0.4, decreases slightly in the early epochs, and then remains relatively stable between 0.38 and 0.41, with minor fluctuations. This trend indicates that the model quickly learns the training data, while performance on the validation set stabilizes after the initial epochs. The observed gap between training and validation loss reflects differences in error on seen vs. unseen data, underscoring the importance of monitoring both metrics to assess model generalization.

Figure 8 presents the calibration plot, which conveys the alignment between predicted probabilities and actual outcomes. Proper calibration ensures the reliability of probabilistic predictions, which are crucial for clinical deployment and risk assessment. The x-axis represents the predicted probability (model confidence) ranging from 0 to 1, while the y-axis shows the empirical accuracy, i.e., the fraction of correct predictions at each confidence level. The dashed diagonal line indicates perfect calibration, where predicted probabilities match observed accuracies exactly (e.g., predictions with 0.7 confidence are correct 70% of the time). The plot line (blue with squares) illustrates the relationship between predicted probabilities and actual accuracies, showing slight underestimation at low confidence levels and modest overestimation at moderate to high confidence levels. The Expected Calibration Error (ECE) of 0.156 quantifies this deviation; lower values indicate better calibration. The reliability diagram is constructed using equal-width confidence bins. For each bin, empirical accuracy is computed as the fraction of correct predictions. The Expected Calibration Error (ECE) is calculated as the weighted average of the absolute difference between confidence and accuracy across all bins.

The final phase focused on interpretability, uncertainty quantification, and clinician-in-the-loop validation to ensure transparency and clinical trustworthiness. Three complementary analyses, Grad-CAM visualization, Monte Carlo Dropout uncertainty estimation, and t-SNE feature embedding, were performed on the trained EfficientOvaNet model. Grad-CAM heatmaps, derived from the final convolutional block, localize salient regions that influence model predictions. Overlays confirmed that EfficientOvaNet consistently emphasized tumor regions rather than background noise. For uncertainty estimation, MC Dropout was re-enabled during inference to generate multiple stochastic forward passes, yielding mean confidence and variance-based uncertainty scores. Misclassified samples exhibited higher uncertainty, suggesting that uncertainty maps could guide radiologists toward cases requiring manual verification.

Figure 9 illustrates predictive uncertainty using MC Dropout. It shows the predictive uncertainty distribution using MC Dropout. Contrary to expectations, both correct and incorrect predictions exhibit near-zero variance, indicating that the MC Dropout implementation did not produce meaningful uncertainty estimates. This suggests limitations in using MC Dropout for uncertainty quantification in our architecture, as the model exhibits uniformly low variance across prediction accuracies. This represents a challenge for identifying ambiguous cases that would benefit from clinical review.

Figure 10 shows t-SNE visualization of high-dimensional feature embeddings. It demonstrates class separability in latent space, providing interpretability into how the model differentiates between tumor categories. Dense clusters indicate strong class discrimination, while overlaps identify challenging cases. In this, the data points are grouped into five tumor classes: class 0, class 1, class 2, class 3, and class 5, represented by blue, orange, green, purple, and red, respectively. The scatter plot displays these classes in a two-dimensional plane, revealing distinct clusters corresponding to different tumor categories. Dense groupings of the red and green classes in the left and bottom regions indicate strong class separability. In contrast, the orange and blue classes appear more dispersed toward the center and right, with slight overlaps suggesting challenging classification regions.

To evaluate the effectiveness of the proposed EfficientOvaNet framework, we compared its performance with several baseline models reported in the literature on the MMOTU dataset. Table 3 summarizes the performance metrics, including accuracy, Dice coefficient, Intersection over Union (IoU), and F1-score. Previous studies primarily focused on segmentation and classification using U-Net variants, DeepLabV3+, PSPNet, DANet, and other deep learning architectures. For instance, Pham et al. (15) achieved a maximum accuracy of 71.65% using DANet on multi-class ovarian tumor segmentation, and PCU-Net (16) improved IoU and Dice by 4.23 and 2.99%, respectively, over the traditional U-Net. Deepika and Chitra (17) demonstrated that the U-Net with BCE-IoU loss achieved a Dice coefficient of 0.8234 and pixel-wise accuracy of 91.5%. The proposed EfficientOvaNet framework, which leverages a dual-branch EfficientNet-B3 architecture with ROI and global contextual features, outperforms these baseline models, achieving a mean accuracy of 91.9%, F1-score of 91.9%, and AUC of 0.98. This improvement can be attributed to the integration of global and local features, advanced preprocessing, data augmentation, class-imbalance control via weighted Focal Loss, and the use of explainable AI methods (Grad-CAM, Monte Carlo Dropout, and t-SNE) to enhance clinical interpretability. EfficientOvaNet demonstrates superior performance compared to prior work, indicating its potential to improve diagnostic accuracy and enable timely intervention in ovarian cancer management.