数据集获取

git clone https://github.com/aparnabg/CNN-Architectures-on-Breast-Ultrasound-Images-Dataset.git

curl -L -o repo.zip https://github.com/aparnabg/CNN-Architectures-on-Breast-Ultrasound-Images-Dataset/archive/refs/heads/main.zip
unzip repo.zip

源站 README 摘录（使用方式）

CNN-Architectures-on-Breast-Ultrasound-Images-Dataset

Comparative analysis of CNN architectures on Breast Ultrasound Images Dataset. Explored 9 CNN models for classifying images into normal, benign, malignant categories. MobileNet and Inception-ResNet-V2 emerged as top performers with 83% and 80% accuracy, respectively. Insights aid in early breast cancer detection.

Introduction

We utilized a publicly available dataset, “Dataset of breast ultrasound images,” for this study. This dataset comprises breast ultrasound images from 600 female patients aged 25 to 75, categorized into normal, benign, and malignant classes. The dataset consists of 780 images with an average resolution of 500 × 500 pixels. Our objective was to explore the efficacy of various convolutional neural network (CNN) architectures for classifying these ultrasound images into the aforementioned categories.

Dataset Characteristics

• Subjects: 600 female patients aged 25 to 75.
• Image Categories: Normal (2), Benign (0), Malignant (1).
• Image Count: 780 images (Benign: 487, Malignant: 210, Normal: 133).
• Image Resolution: Average size of 500 × 500 pixels.
• Data Collection: Images collected using LOGIQ E9 and LOGIQ E9 Agile ultrasound systems.
• Preprocessing: Removed duplicates, corrected annotations, and converted DICOM to PNG.
• Ground Truth Annotation: Accurate delineation of image boundaries.

Data Augmentation

To address class imbalance, we employed data augmentation techniques including rotation, shifting, zooming, shearing, and flipping. This approach generated additional samples, balancing the dataset and enhancing model robustness.

Implemented CNN Architectures

We implemented the following CNN architectures:

• EfficientNet
• Inception-ResNet-V2
• ResNet50
• MobileNet
• VGG16
• Depth-based CNN
• AlexNet
• VGG19
• PolyNet

Results and Conclusion

The accuracy of the implemented models ranged from 15% to 83%.

Top Performers:

• MobileNet: 83% accuracy. Efficient feature extraction due to its lightweight architecture.
• Inception-ResNet-V2: 80% accuracy. Impressive due to its innovative combination of Inception and ResNet architectures.

Other Observations:

• VGG16 and VGG19: Achieved high accuracies but exhibited higher losses, suggesting potential overfitting.
• Depth-based CNN and AlexNet: Showed lower accuracies.
• **

数据加载示例（图像类）

from PIL import Image
import glob, os

files = (glob.glob(os.path.join(path, "**", "*.png"), recursive=True)
       + glob.glob(os.path.join(path, "**", "*.jpg"), recursive=True)
       + glob.glob(os.path.join(path, "**", "*.tif"), recursive=True))
print("图像文件数:", len(files))
img = Image.open(files[0]); print("尺寸/模式:", img.size, img.mode)

# torchvision Dataset 方式:
# from torchvision import datasets
# ds = datasets.ImageFolder(path)  # 要求 子目录=类别

目录组织与标注格式以源站说明和下载后实际文件为准。

完整仓库：github.com/aparnabg/CNN-Architectures-on-Breast-Ultrasound-Images-Dataset