Introduction

Breast cancer is a malignant disease characterized by uncontrolled proliferation of epithelial cells in breast tissue and remains a major global health challenge [1–3]. Its etiology is multifactorial, involving genetic susceptibility, hormonal influences, and lifestyle and environmental exposures [4–6]. Because clinical outcomes are closely linked to stage at diagnosis, timely and accurate detection remains central to effective management, which has sustained interest in noninvasive imaging-based diagnostic pathways [7, 8].

Mammography is the established standard for population-level screening; however, its diagnostic sensitivity decreases in women with dense fibroglandular tissue and can be lower in younger populations [9–12]. This limitation motivates the use of complementary imaging modalities to reduce missed findings and improve diagnostic confidence [13–15]. Ultrasound is widely used as an adjunct because it is accessible, does not involve ionizing radiation, and supports dynamic assessment in real time [16]. It is particularly valuable for clarifying equivocal mammographic findings and for evaluation in dense breasts [17]. At the same time, ultrasound interpretation can be operator dependent, contributing to interobserver variability and inconsistent reporting across settings [18–20]. These challenges strengthen the case for analysis approaches that are objective, repeatable, and easier to standardize.

In parallel, artificial intelligence-driven image analysis has expanded rapidly across medical imaging, including oncologic applications [21, 22]. Recent work suggests that deep learning systems can support classification, detection, and segmentation tasks, with the potential to reduce variability in routine interpretation when appropriately validated [23–25]. In breast ultrasound specifically, the central question is not only whether automated models can achieve high accuracy, but also whether they do so in a way that is practical for real-world use.

Architecturally, Convolutional Neural Networks (CNNs) have been the dominant approach because they learn robust local features efficiently [26–28]. Vision Transformer (ViT) models have emerged as an alternative by modeling broader contextual relationships, which may be advantageous when diagnostically relevant cues extend beyond localized textures [29, 30]. At the same time, transformer-based approaches may require higher computational resources and can be more sensitive to dataset scale and heterogeneity, which are common constraints in medical imaging research [31].

Recent studies in breast ultrasound have increasingly explored hierarchical transformer designs and multimodal strategies to combine global context with localized feature extraction [32]. For example, multimodal transformer-based fusion has been evaluated for malignancy prediction in retrospective multicenter settings [33], and transformer-driven end-to-end frameworks have been reported for multicenter molecular subtype classification using multimodal ultrasound-derived information [34]. These advances indicate active progress, yet direct benchmarking studies that compare modern transformer models with widely used CNN baselines while also reporting computational efficiency remain relatively limited.

Motivated by this gap, the present study provides a focused comparative evaluation of two transformer-based models, Swin Transformer Base and DeiT Base, against two established CNN architectures, InceptionV3 and MobileNetV3 Large, for three-class breast ultrasound classification. Beyond predictive performance, we explicitly consider efficiency-related indicators to clarify the tradeoff between accuracy and computational demand, supporting more informed model selection for different deployment constraints.

Materials and methods

Dataset

We used the publicly available Breast Ultrasound Images (BUSI) dataset distributed via Kaggle [35]. BUSI comprises 780 ultrasound images labeled into three categories: benign, malignant, and normal. This three-class structure supports a direct evaluation of how well different architectures separate lesion types while also distinguishing normal tissue patterns.

Images were separated using a stratified split of 70 percent for training (n = 545), 15 percent for validation (n = 115), and 15 percent for testing (n = 120). Although K-fold cross-validation is a viable alternative, a fixed hold-out split was preferred here to ensure an identical and completely independent test environment for comparing diverse CNN and Transformer paradigms. Stratification was applied to preserve the original class proportions across subsets and to reduce the risk of majority class dominance during optimization, given the smaller size of the normal category (n = 133) relative to benign (n = 437). The validation subset was used for model selection and early stopping, while the test subset was held out and used only for the final comparison across models. A full breakdown is provided in Table 1.

Table 1. Distribution of BUSI images across training, validation, and test subsets.

Classes	Train (70%)	Validation (15%)	Test (15%)	Total (100%)
Benign	305	65	67	437
Malignant	147	31	32	210
Normal	93	19	21	133
Total	545	115	120	780

Display full size

BUSI: Breast Ultrasound Images.

Figure 1 provides representative examples from each class to illustrate the visual variability encountered in BUSI. In general, benign lesions often present more regular margins and relatively homogeneous echotexture, whereas malignant lesions more frequently show irregular borders and heterogeneous internal appearance. Normal examples are included to reflect typical background parenchymal patterns. The figure also highlights common ultrasound characteristics such as speckle and limited contrast, which contribute to the difficulty of the classification task.

Display full size

Figure 1. Representative BUSI examples for malignant, benign, and normal categories. BUSI: Breast Ultrasound Images.

Results

The comparative results are summarized in Table 2, which reports predictive performance together with model size and computational complexity. We further examine error patterns for the best-performing model using the confusion matrix in Figure 2 and provide qualitative visual explanations using Gradient Weighted Class Activation Mapping (Grad-CAM) in Figures 3 and 4.

Display full size

Figure 2. Confusion matrix for Swin Transformer Base on the independent test set.

Display full size

Figure 3. Grad-CAM examples for Swin Transformer Base, showing representative correct and misclassified predictions.

Display full size

Figure 4. Representative correct and incorrect Swin Transformer Base predictions with prediction confidence values.

Swin Transformer Base achieved the highest overall performance, reaching an accuracy of 0.9167 and an F1-score of 0.8981. This gain came with a substantial computational footprint, including 86.75M parameters and 30.3375 Giga Floating-Point Operations (GFLOPs). In contrast, MobileNetV3 Large reached an accuracy of 0.8583 with 4.21M parameters and 0.4307 GFLOPs. Relative to Swin, MobileNetV3 Large uses approximately 20.61 times fewer parameters and approximately 70.44 times fewer GFLOPs, making the efficiency contrast more pronounced than the accuracy gap.

DeiT Base ranked second by accuracy (0.8750) and achieved an F1-score of 0.8555. Its parameter count (85.80M) is comparable to Swin, while its GFLOPs are slightly higher (33.6955). The CNN baselines placed behind the transformer models in accuracy, with InceptionV3 and MobileNetV3 Large both reaching 0.8583. However, their resource profiles differed markedly. InceptionV3 required 21.79M parameters and 5.6719 GFLOPs, whereas MobileNetV3 Large matched the accuracy with approximately 5.18 times fewer parameters and approximately 13.17 times fewer GFLOPs. These results highlight a consistent pattern: higher accuracy was associated with higher computational demand, while the most compact model provided the strongest efficiency profile.

From an application perspective, the practical choice depends on the deployment context. If the primary aim is maximizing performance under ample compute, a model such as Swin Transformer Base may be preferred. When compute budgets are limited or latency constraints are strict, MobileNetV3 Large offers a favorable accuracy to cost balance. These workflow-oriented interpretations should be treated as provisional until supported by direct measurements of inference latency, memory use, and throughput on the target clinical hardware.

Discussion

The results in Table 2 indicate a consistent tradeoff between predictive performance and computational demand when comparing ViT models with CNN baselines for breast ultrasound classification. Swin Transformer Base achieved the strongest overall performance, with an accuracy of 0.9167, supported by a high precision of 0.9409 and an F1-score of 0.8981. This pattern is compatible with the notion that hierarchical attention, implemented through shifted windows, can capture both local and broader contextual cues that are relevant in ultrasound images. At the same time, the computational cost is substantial, with 86.75M parameters and 30.3375 GFLOPs.

In contrast, MobileNetV3 Large delivered an accuracy of 0.8583 at a markedly lower computational footprint, using 4.21M parameters and 0.4307 GFLOPs, representing approximately 20.61 times fewer parameters and approximately 70.44 times fewer GFLOPs than Swin Transformer Base. This gap suggests that a large fraction of the performance obtained by the strongest model can be retained with far lower computational requirements, which may be important when inference latency, memory limits, or energy constraints are central. These deployment-oriented interpretations should remain cautious until confirmed with direct measurements of inference time, throughput, and memory use on the target hardware.

DeiT Base ranked second with an accuracy of 0.8750 and an F1 score of 0.8555, while maintaining a parameter count comparable to Swin Transformer Base (85.80M). The two CNN baselines, InceptionV3 and MobileNetV3 Large, reached the same accuracy value (0.8583) but differed substantially in cost. InceptionV3 required 21.79M parameters and 5.6719 GFLOPs, whereas MobileNetV3 Large achieved the same accuracy with approximately 5.18 times fewer parameters and approximately 13.17 times fewer GFLOPs. This contrast reinforces that architecture choice can materially change operational burden even when headline accuracy is similar.

Limitations should be considered when interpreting these findings. First, the study relies on a single public dataset of 780 images, which may not capture the full variability seen across scanners, acquisition protocols, and patient populations. Second, the class distribution is imbalanced, with fewer normal cases (n = 133) than benign cases (n = 437), which can influence model behavior across classes. Third, the dataset redistribution does not consistently provide patient identifiers, so the split is performed at the image level, and patient-level overlap across subsets cannot be fully excluded. Future work should prioritize external validation on multi-center cohorts, patient-level splitting when identifiers are available, and reporting that supports clinical translation, such as confidence intervals, calibration assessment, and a direct comparison with radiologist performance or established commercial computer-aided diagnosis systems. If the intended use case prioritizes specific error types, such as reducing false negatives for malignant lesions, cost-sensitive objectives, or class-balanced training strategies can be evaluated in a targeted way.

To summarize the overall impact and findings of this comparative evaluation. This study compared two ViT models (Swin Transformer Base and DeiT Base) and two CNN models (InceptionV3 and MobileNetV3 Large) for three-class breast ultrasound classification, emphasizing both performance and computational efficiency. Swin Transformer Base achieved the highest performance, reaching an accuracy of 0.9167 and an F1 score 0.8981, but with high computational demand. MobileNetV3 Large achieved an accuracy of 0.8583 with a substantially lower footprint, supporting its suitability when efficiency constraints are dominant. The results suggest that architecture selection should be aligned with the practical requirements of the intended workflow, and that stronger claims about deployment readiness should be supported by external validation and hardware-specific inference benchmarking.