A Hybrid Mobile Net-Active Learning Framework for Breast Cancer Detection Using Ultrasound Images

Specifically, the goal will be to increase the precision of image categorization for breast cancer through an optimal deep learning-based methodology, which can overcome hurdles such as data imbalance and hyper parameter tuning to improve performance. At a high level, the process consists of dataset aggregation, image preprocessing and a CNN-based classification pipeline, which is further fine-tuned using a bio-inspired algorithm. This process starts by merging two datasets (from Kaggle and our own). They are then merged to create a diverse and exhaustive training, validation and testing data pool. Three classes—normal, benign, and malignant—are applied to the photographs. In cases of class imbalance, where certain categories (for example, verbal scores) are overrepresented and others aren't, data balancing techniques are applied to ensure that all categories have sufficient training samples.

Image preprocessing steps are conducted to improve image quality and consistency. Prior to being fed into the CNN model, the pictures undergo pre-processing, which includes normalization and resizing to a standard dimension. Besides, Data augmentation methods like rotation are used to enhance the dataset, including shifting, zooming, and flipping, which makes it much larger than its original size, thereby increasing the model's robustness towards variations in input data. CNN Architecture for Three-Class Breast Cancer Classification. This design includes max pooling layers for down-sampling, convolutional layers for feature extraction, batch normalization for stabilization and dense layers for classification. We compile the model using Adam with categorical cross-entropy loss, and to avoid over fitting, we use early stopping.

3.png

F1 score, precision, accuracy, confusion matrices, recall, and other performance indicators are among the various calculations that comprise the approach. These are the metrics we use to cross-check the effectiveness of our model on the validation dataset. The models are evaluated graphically using learning curves and confusion matrices to gain insights into their strengths and weaknesses. Finally, the methodology saves all results into CSV files for reproducibility and further analysis. Finally, this end-to-end process incorporates all the steps of data preparation, augmentation, and preprocessing, as well as building and training the models, to optimize for best accuracy and present a final solution that has proven efficiency in the classification of breast cancer.

The following subsection explains the step-by-step approach we propose, as illustrated in Figure 1, from dataset preparation through to CNN model evaluation after optimisation.

3.1 Dataset overview

3.1.1 Kaggle dataset

The "Ultrasound Breast Images for Breast Cancer" dataset, obtained from Kaggle, was utilized in this investigation. Ultrasound pictures of benign and malignant breast cancers make up this dataset, which is essential for tasks involving the identification and classification of malignancy. The dataset contains sufficient images for reliable training and validation, thanks to augmentation techniques such as rotation and sharpening.

• Class malignant: 4042 images

• Class benign: 4074 images

This is a popular practice dataset for deep learning and machine learning problems because of its size and diversity. An example of malignant and benign ultrasound pictures from this dataset is displayed in Figure 2. The model describes two groups based on the images they mention that show different forms of breast tissue.

Detailed variability of the dataset, as well as augmentation methods used, make this dataset good enough to train deep learning models and balance the representation of such classes. In this work, the Kaggle dataset was used as a supplementary source to strengthen class representation and improve generalization.

3.1.2 Proposed collection dataset

The three primary types of breast ultrasound pictures in the proposed collection dataset are benign, malignant, and normal. The dataset was sourced from an oncology teaching hospital in Iraq, and the preprocessed photos in this dataset are used to enhance machine learning and deep learning methods to detect and classify breast cancer, which consists of:

• Class malignant: 429 images

• Class benign: 289 images

• Class normal: 72 images

The dataset comprises a diverse range of ultrasound images representing various breast tissue conditions. There are three classes: (a) normal, (b) malignant, and (c) benign. This dataset is clinically valuable because it introduces normal breast tissue cases, which are not present in the Kaggle dataset, and reflects imaging variations encountered in real-world diagnostic settings.

Figure 3 shows an example ultrasound image from the normal, malignant, and benign classes. It demonstrates the complexity of visual features typically found in such cases, which are critical for classification.

The size distribution of pictures in the benign, malignant, and normal classifications is shown in Figure 4. The width and height distributions for each class are displayed individually in the histograms, which demonstrate variation in picture dimensions.

Figure 2. Example images per class: Benign (Top) and malignant (Bottom)

1.png

Figure 3. An example from the dataset of proposed collection. the images represent a benign, malignant, and normal tumor, illustrating key diagnostic features

4.png

Figure 4. Image size distribution for malignant, benign, and normal classes. each subplot represents the width and height frequency distribution of images for a specific class

Figure 6 demonstrates the balanced distribution of the dataset after augmentation and over-sampling. Now, all three classes (malignant, benign and normal) have the same number of images, preventing model bias and allowing for effective training.

5.png

Figure 5. Class distribution of the proposed collection dataset before balancing. Malignant images dominate the dataset, followed by benign and normal classes

6.png

Figure 6. Balanced class distribution after data augmentation. Each class has an equal number of images, ensuring fairness in model training

3.1.3 Merged dataset

To construct a comprehensive dataset, the two sources were merged into three categories: malignant, benign and normal. The Iraqi dataset contributed unique normal images, while both the Kaggle and Iraqi datasets contributed benign and malignant cases. After merging, oversampling and augmentation techniques (rotation, flipping, shifting, zooming, and sharpening) were applied to balance the classes. The final balanced dataset consisted of 4,471 images per class (normal, benign, and malignant), ensuring equal representation across categories.

• Class normal: 4471 images

• Class benign: 4471 images

• Class malignant: 4471 images

This single dataset facilitates incremental learning from images of each class, enabling equitable testing and training of deep learning models. It ensures that the dataset is well-balanced, allowing the model to learn from a diverse range of samples. Hence, increasing the classification performance whilst decreasing over fitting.

3.1.4 Data splitting strategy

To establish a reliable evaluation protocol and minimize the risk of over fitting, the merged dataset was divided into distinct training, validation, and test subsets. A stratified patient-level splitting strategy was employed to ensure that all images from a single patient were confined to a single subset. This precaution prevents the leakage of patient-specific patterns across training and evaluation phases, which could otherwise artificially inflate performance metrics. The division followed a 70/15/15 ratio, allocating the majority of images to the training subset, while ensuring that the validation and test subsets remained sufficiently large and balanced for meaningful evaluation. Stratification preserved the proportional representation of the three classes (normal, benign, malignant) across all subsets, thereby maintaining class consistency and reducing bias during model development.

The validation subset was utilized for hyper parameter tuning, optimization, and early stopping to prevent over fitting during training. The test subset, held out entirely from the training and tuning process, was reserved exclusively for the final performance evaluation. This design ensured that the F1 scores, recall, precision, and accuracy reported in the study accurately reflected the actual generalization capacity of the proposed framework, rather than relying on the memorization of redundant patient-level features. By maintaining patient-level independence across subsets and applying a systematic splitting strategy, the evaluation process provided a fair and reproducible assessment of the model's diagnostic performance in breast cancer ultrasound image classification.

3.2 Convolution neural network architecture

Conv Net Architecture: A convolutional neural network (CNN) architecture was created specifically with the aim of dividing breast ultrasonography into three categories: benign, malignant, and normal. It commences with an input layer that receives 224×224 pictures with three color channels. The input data is normalized to make it similar for efficient training. The network is composed of a sequential structure, with convolutional layers serving as the primary building block for feature extraction. A pooling layer is employed for dimensionality reduction, and batch normalization is used within the convolutional layers to stabilize and accelerate the training process.

In the first layer, we incorporate 32 convolutional filters with a 3×3 kernel size and introduce non-linearity using the ReLU activation function. A max-pooling layer follows, which reduces computation while preserving significant characteristics by down-sampling the spatial resolution by calculating the maximum value in each local 2×2 region. A batch normalization layer is added after the pooling procedure to normalize the activations and stabilize the training process. In the following layers, this process is repeated with increasing the filters to 64 and 128, respectively, for the second and third convolutional layers. With each layer we build on top of the previous layers, we extract higher-level and more complicated features that allow us to distinguish very subtle differences in the input images.

A fully connected dense layer with 256 neurons receives the feature maps after they have been flattened into a one-dimensional vector following the convolutional and pooling processes. The ReLU activation function is employed by this thick layer to identify intricate links and patterns in the data. A dropout layer, which randomly deactivates 50% of the neurons during training, is incorporated to prevent overfitting. Three neurons representing the three output classes (benign, normal and malignant), make up the last dense layer. The SoftMax activation function is applied in this layer to translate the network’s outputs into probabilities that accurately reflect the likelihood of each class and ensure that they add up to one.

The Adam optimizer, which dynamically modifies the learning rate during training for effective convergence, is used to build the network. Since it is a multi-class issue, we choose categorical cross-entropy as our loss function. Accuracy is also supported as a performance metric for both validation and training. Additionally, we used early stopping based on the validation loss, which stops training if the validation loss does not improve after five epochs, in order to prevent overfitting and achieve a robust performance.

Table 2 presents the configuration and parameters of the proposed CNN architecture.

Table 2. Convolutional neural network architecture parameters

The augmented data is then used to train the CNN by applying transformations such as rotation, width and height shifts, shearing, zooming, and horizontal flipping. Such data augmentation enhances the model's generalization by exposing it to a broader range of image variations. It updates weights iteratively to minimize the loss, while performing validation on a separate validation dataset to verify its learning progress. A confusion matrix, which provides additional information about the contrast between predicted classes and their actual values, and several quantitative metrics (accuracy, precision, recall, and F1 score) are used to summarize the feedback gathered from the previous design. This aspect enables a powerful yet efficient CNN architecture, well-suited for classifying breast ultrasound images. The outline of the convolutional neural network (CNN) applied to breast ultrasound images classification is outlined in Table 1. A human-readable description of each layer and its parameters, along with their values, describes the structure and configuration of the network.

3.3 Transfer learning

Transfer learning was employed through the adoption of pre-trained convolutional neural network architectures, namely MobileNet, VGG16, VGG19, and Xception. Each model was initialized with ImageNet weights, and a two-stage fine-tuning strategy was adopted to optimize both computational efficiency and task-specific adaptation for breast ultrasound classification. In the first stage, the convolutional base of each architecture was frozen, and only the newly introduced classification head was trained. This design enabled the networks to retain their pre-acquired low-level feature representations while progressively adapting their higher-level abstract features to the domain of ultrasound images. Such an approach mitigates the limitations of small and imbalanced medical datasets, ensuring that knowledge from large-scale natural image datasets is effectively leveraged while avoiding overfitting at the early stages of training.

In the second stage, selective fine-tuning was performed to further refine the feature representations relevant to the task. For MobileNet, the final depthwise separable convolutional block was unfrozen; for VGG16 and VGG19, the last two convolutional blocks were retrained; and for Xception, the final 36 layers corresponding to deeper separable convolutions were fine-tuned. To ensure stable convergence and avoid catastrophic forgetting, a reduced learning rate of 1×10^-5 was employed during this phase, while earlier layers remained frozen to preserve generalizable features. All models were trained using the Adam optimizer with categorical cross-entropy loss, and early stopping (patience =8 epochs) was employed to mitigate over fitting risks. This two-phase training strategy allowed the models to gradually transition from generic feature extraction toward ultrasound-specific discriminative learning, yielding more robust and clinically meaningful representations.

The classification heads of the networks were tailored to the task requirements. MobileNet accepted 224×224×3 inputs, followed by a Global Average Pooling layer, a dense layer with 256 neurons activated by ReLU, and a dropout of 0.5 to enhance regularization. VGG16 and VGG19 maintained their convolutional backbones but incorporated a flatten operation, a 256-neuron dense layer with ReLU, and a dropout layer. Xception, which required 299×299×3 inputs, employed a Global Average Pooling layer followed by a dense layer of 256 neurons and a dropout of 0.5. In all architectures, the final classification layer contained three neurons with softmax activation corresponding to the classes normal, benign, and malignant. As shown in Table 3, the MobileNet, VGG16, VGG19, and Xception models differ in input size, pre-training datasets, and architecture-specific configurations. To strengthen generalization, extensive data augmentation was applied, including rotations, shifts, shearing, zooming, and flipping, thereby simulating realistic variations in ultrasound acquisition. Comparative results demonstrated that MobileNet achieved the highest efficiency in balancing accuracy and computational cost. VGG16 and VGG19 delivered strong classification through hierarchical feature extraction, and Xception proved effective in modeling high-level representations despite albeit with slightly lower accuracy. These findings substantiate selective fine-tuning of transfer learning architectures as a robust methodology for medical image classification, particularly in domains constrained by limited annotated data.

The Snake Optimizer (SO) is a bio-inspired metaheuristic algorithm based on the mating and feeding behaviors of snakes. The algorithm simulates the relationship between the subpopulations of males and females in response to environmental influences, such as food and temperature. SO provides a balance between diversification and intensification in the search process through the alternation of global and local exploitation. The mechanism mitigates the risks of early convergence and local optimum traps, which are common issues in high-dimensional optimization problems, particularly when employing deep learning models.

SO has been selected for this study because it offers comparative advantages over existing optimization methods. Although conceptually simple, Grid Search and Random Search involve searching the space exhaustively or randomly, which is computationally infeasible at large-scale hyper parameter tuning. Bayesian Optimization improves efficiency because it models the search as a probabilistic process; however, in highly irregular or non-smooth landscapes, its performance is poor. Although both Genetic Algorithms and Particle Swarm optimization are efficient, both are often slower to converge and are also sensitive to parameter settings. Benchmark studies have shown that SO is converging faster, less computationally costly, and more robust in adopting a balance between exploration and exploitation. This benefit was further supported by empirical analysis: SO achieved better classification accuracy with fewer iterations than Random Search using the same CNN parameter space, which proved its usefulness as a hyper parameter optimization method in medical imaging tasks.

3.4.1 Initialization

A population of potential solutions, or a group of candidate solutions, is where the optimization process begins. These are first scattered throughout the search space, which is determined by the search parameters.

$X_i=X_{\min }+r \times\left(X_{\max }-X_{\min }\right)$                (1)

where, r is a random number in [0, 1], X_i$~$is the location of the ith person, and X_min, X_max$~$are lower/upper bounds on the search space.

3.4.2 Group division

Population consists of two sexes, male (N_m) and female (N_f), described here.

$N_m=[N / 2]$            (2)

$N_f=N-N_m$             (3)

3.4.3 Exploration phase

If food quantity is not available, snakes reach the exploration state, where they explore different locations in the solution space:

$\begin{aligned} X(t+1)=X \operatorname{rand}, & m \pm c 2 \cdot A m \cdot((X \max -X \min ) \cdot r+X \min )\end{aligned}$                (4)

where, r is a random number, A_m$~$denotes the male’s capacity to find food, and X^(t+1) is the updated position of the ith male.

3.4.4 Exploitation phase

In the presence of food, snakes transition to exploitation, focusing on the most promising areas:

$X^{(t+1)}=X_{\mathrm{food}} \pm c_3 \cdot \operatorname{Temp} \cdot r \cdot\left(X_{\mathrm{food}}-X^{(t)}\right)$                (5)

where, X_food represents the position of the food, Temp is the current temperature, and c₃ is a constant.

3.4.5 Mating and fight mode

If the environmental conditions are optimal, snakes engage in mating or fighting modes. The fight mode updates positions as follows:

$X^{(t+1)}=X^{(t)}+c_3 \cdot F_M \cdot r \cdot\left(Q \cdot X_{\mathrm{best}, f}-X^{(t)}\right)$            (6)

$X^{(t+1)}=X^{(t)}+c_3 \cdot F_F \cdot r \cdot\left(Q \cdot X_{\text {best }, m}-X^{(t)}\right)$                  (7)

where the fighting skills of males and females are denoted by F_M$~$and F_F, respectively. Here, F_M$~$and F_F, represent the fighting ability coefficients for male and female individuals, respectively.

In the mating mode, positions are updated based on mutual interaction:

$X^{(t+1)}=X^{(t)}+c_3 \cdot M_m \cdot r \cdot\left(Q \cdot X^{(t)}-X^{(t)}\right)$                (8)

$X^{(t+1)}=X^{(t)}+c_3 \cdot M_f \cdot r \cdot\left(Q \cdot X^{(t)}-X^{(t)}\right)$                   (9)

where, male and female mating capacities are indicated by M_m$~$and M_f.

3.4.6 Termination

Until a termination criterion (a convergence threshold is satisfied or the maximum number of iterations is reached), the algorithm keeps running repeatedly.

The Snake Optimizer algorithm demonstrates strong exploration and exploitation capabilities, and thus is suited for solving a wide range of optimization problems effectively.

Through the optimization process using the Snake Optimizer Algorithm, the best set of hyper parameters for the Convolutional Neural Network (CNN) was identified. The optimal learning rate was determined to be 0.0001, which ensures gradual and stable updates to the model weights during training. A dropout rate of 0.2 was selected, effectively mitigating overfitting by randomly deactivating 20% of the neurons during training. The Adam optimizer was also selected because it combines the benefits of momentum and adjustable learning rates, which improve model performance and speed up convergence.

The best score achieved using these parameters on the validation dataset was 0.8375, indicating a balanced and effective configuration for the CNN architecture. These parameters will be used in the final training phase to ensure optimal performance and generalization of the model.

3.5 Active learning

To enhance model efficiency under limited labeled data, we employed a pool-based Active Learning (AL) strategy in conjunction with MobileNet. In this setting, the full dataset D was divided into a small initial labeled set $\mathcal{L}_0$ and a large unlabeled pool $u_0$. Specifically, $\mathcal{L}_0$ consisted of 600 stratified images (200 per class: normal, benign, malignant) randomly selected to maintain balance across categories, while the remainder of the dataset formed $u_0$.

At each AL round t , the MobileNet model was trained on the current labeled set $\mathcal{L}_t$ and validated on a held-out subset. The trained model then produced probability distributions $\mathrm{p}(\mathrm{x})$ for all samples $\mathrm{x} \in u_t$. To quantify prediction uncertainty, we calculated predictive entropy:

$\mathrm{H}(\mathrm{x})=-\sum_{\mathrm{k}=1}^{\mathrm{K}} p_k(x) \cdot \log p_k(x), \mathrm{K}=3$

Samples with the highest entropy were considered the most uncertain and thus, the most informative for labeling. At each iteration, a batch of unlabeled samples with the highest entropy scores was selected and annotated by a simulated oracle. These samples were then added to the labeled set $\mathcal{L}_{t+1}$ and removed from the pool $u_t$.

The MobileNet model was retrained at every iteration using warm-starting from the previously saved weights, with early stopping to prevent overfitting. This cycle was repeated for multiple AL rounds until the label budget was reached or validation performance plateaued.

By progressively querying the most uncertain samples, the AL framework ensured that the model concentrated on the most challenging instances, leading to improved accuracy, F1-score and recall compared to training with a static labeled set. This demonstrates the advantage of entropy-based Active Learning for efficient breast cancer ultrasound image classification in settings with limited labeled data.

To strengthen the reliability of the reported results, we performed repeated experiments with different random seeds (n=5) and analyzed the outcomes using confidence intervals and effect size statistics. The detailed numerical outcomes are provided in Table 7, which confirms the statistical significance of the observed improvements. For the MobileNet baseline, the mean accuracy was 94.56%±0.36%, with a 95% confidence interval of [94.11%, 95.01%]. In contrast, the proposed MobileNet_AL achieved 96.12%±0.05%, with a 95% confidence interval of [96.06%, 96.18%]. The non-overlapping confidence intervals confirm that the improvement is statistically significant. Furthermore, Cohen’s d was computed at 6.07, indicating a considerable effect size and validating that the observed performance gains are not due to random variation. Similar trends were observed across precision, recall, and F1-score, all showing consistently minor variance and higher mean values for MobileNet_AL compared to the baselines.

Table 7. Summary of experimental results with statistical validation and cross-dataset testing

In addition to repeated measures, we conducted a cross-dataset validation to evaluate the generalization capability of the framework. The Kaggle dataset was used for training while the Iraqi dataset served exclusively as the test set, and vice versa. When trained on Kaggle and tested on Iraqi data, MobileNet_AL maintained an accuracy above 90%, confirming its ability to adapt to unseen clinical images from a different source. Conversely, training on the Iraqi dataset and testing on Kaggle achieved slightly lower accuracy (~88%) because of the Iraqi dataset’s smaller size, but the model still preserved balanced recall across the three classes. These experiments validate that the unique Iraqi dataset enhances real-world diversity and supports the robustness of the proposed model, addressing a limitation of many prior studies that rely solely on public datasets.