Text

Introduction: Breast cancer remains the most prevalent cancer among women, making early detection critical for reducing mortality rates. Ultrasound imaging is a valuable diagnostic tool due to its accessibility and non-invasiveness; however, its accuracy depends heavily on the expertise of the sonographers, which may lead to diagnostic errors [1]. To address this challenge and support clinicians, particularly those with limited experience, a prototype computer-aided diagnosis (CAD) tool is proposed that integrates deep convolutional neural networks (DCNNs) for breast lesion classification and segmentation in ultrasound images.

Methods: The classification model is based on a dual-path convolutional neural network (DPCNN) [2], which utilizes a pre-trained ResNet-50 backbone, leveraging transfer learning to improve performance with limited data. To delineate the contours of the lesions, we implemented a U-Net model by Baccouche et al. [3] for segmentation. The models were trained on a curated dataset comprising 1,120 ultrasound images from 218 patients, covering three BI-RADS categories: B1, B2, and B5. The ground truth masks were generated through a hybrid approach combining manual annotation and Meta AI’s Segment Anything tool. In order to mitigate overfitting, a range of data augmentation techniques such as random rotations (±10%), random zoom (±20%), horizontal mirroring, and combinations of these transformations were applied.

Results: The classification model achieved a sensitivity and positive predictive value (PPV) of 99.4%, thereby affirming its reliability in differentiating between benign and malignant lesions. Moreover, the U-Net segmentation model yielded a Dice coefficient of 97.3%, indicating high spatial accuracy in lesion segmentation. The prototype provides visual and probabilistic feedback through a user-friendly interface, marking the lesion area and displaying the model's confidence scores. However, the prototype's performance on the external BUSI dataset [4] dropped significantly (accuracy: 60%, Dice: 17%), highlighting challenges in model generalizability due to domain-specific differences.

Discussion: A comparison of the prototype's performance with commercially available systems, such as Samsung’s S-Detect tool [5] (sensitivity: 95.8%, specificity: 62–93.8%) and the research-based segmentation model by Baccouche et al. [3] (Dice > 95%), reveals that our prototype demonstrates competitive results despite the limited dataset. Nevertheless, differences in dataset composition and model architecture preclude direct comparison. Identified limitations include the restricted spectrum of BI-RADS categories and a certain degree of similarity between the training and test datasets due to the limited number of patients, which may contribute to optimistic performance estimates.

Conclusion: This study presents a deep learning-based CAD prototype for the analysis of breast ultrasound images, incorporating lesion classification and segmentation. Despite strong internal performance, external validation results emphasize the need for broader datasets and cross-device training to ensure clinical robustness. Future efforts should focus on expanding the dataset to cover the full BI-RADS spectrum and performing evaluations across multiple imaging devices to enhance model generalizability.

The authors declare that they have no competing interests.

The authors declare that a positive ethics committee vote has been obtained.

Literatur

[1] Brunetti N, Calabrese M, Martinoli C, Tagliafico AS. Artificial Intelligence in Breast Ultrasound: From Diagnosis to Prognosis — A Rapid Review. J Ultrasound. 2021;25(1):11–21.
[2] Shahzad A, Mushtaq A, Sabeeh AQ, Ghadi YY, Mushtaq Z, Arif S, et al. Automated Uterine Fibroids Detection in Ultrasound Images Using Deep Convolutional Neural Networks. Healthcare (Basel). 2023;11(6):1–14.
[3] Baccouche A, Garcia-Zapirain B, Castillo Olea C, Elmaghraby AS. Connected-UNets: a deep learning architecture for breast mass segmentation. NPJ breast cancer. 2021;7(1):151.
[4] Al-Dhabyani W, Gomaa M, Khaled H, Fahmy A. Dataset of breast ultrasound images. Data brief. 2020;28:104863.
[5] Xia Q, Cheng Y, Hu J, Huang J, Yu Y, Xie H, et al. Differential diagnosis of breast cancer assisted by S-Detect artificial intelligence system. Math Biosci Eng. 2021;18(3):3680–89.