Twum Frimpong, Ahiable Charlyne Carol Eyram, Oppong Stephen Opoku, Banning Linda, Owusu-Agyemang Kwabena
Department of Computer Science, Kwame Nkrumah University of Science and Technology, Kumasi, Ghana.
PLOS Digit Health. 2025 Jun 16;4(6):e0000907. doi: 10.1371/journal.pdig.0000907. eCollection 2025 Jun.
Breast cancer remains a critical global health concern, affecting countless lives worldwide. Early and accurate detection plays a vital role in improving patient outcomes. The challenge lies with the limitations of traditional diagnostic methods in terms of accuracy. This study proposes a novel model based on the four pretrained deep learning models, Mobilenetv2, Inceptionv3, ResNet50, and VGG16, which were also used as feature extractors and fed on multiple supervised learning models using the BUSI dataset. Mobiletnetv2, inceptionv3, ResNet50 and VGG16 achieved an accuracy of 85.6%, 90.8%, 89.7% and 88.06%, respectively, with Logistic Regression and Light Gradient Boosting Machine being the best performing classifiers. Using transfer learning, the top layers of the model were frozen, and additional layers were added. A GlobalAveragePooling2D layer was employed to reduce spatial dimensions of the input image. After training and testing based on the accuracy, ResNet50 performed the best with 95.5%, followed by Inceptionv3 92.5%, VGG16 86.5% and lastly Mobilenetv2 84%.
乳腺癌仍然是一个关键的全球健康问题，影响着世界各地无数人的生命。早期准确检测在改善患者预后方面起着至关重要的作用。挑战在于传统诊断方法在准确性方面存在局限性。本研究提出了一种基于四个预训练深度学习模型（MobileNetv2、Inceptionv3、ResNet50和VGG16）的新型模型，这些模型还被用作特征提取器，并使用BUSI数据集输入到多个监督学习模型中。MobileNetv2、Inceptionv3、ResNet50和VGG16的准确率分别达到了85.6%、90.8%、89.7%和88.06%，其中逻辑回归和轻梯度提升机是表现最佳的分类器。使用迁移学习，模型的顶层被冻结，并添加了额外的层。采用全局平均池化2D层来减少输入图像的空间维度。在基于准确率进行训练和测试后，ResNet50表现最佳，为95.5%，其次是Inceptionv3为92.5%，VGG16为86.5%，最后是MobileNetv2为84%。
