数据集获取

git clone https://github.com/VahidMohammadZadehEivaghi/Skin-Cancer-Detection.git

curl -L -o repo.zip https://github.com/VahidMohammadZadehEivaghi/Skin-Cancer-Detection/archive/refs/heads/main.zip
unzip repo.zip

源站 README 摘录（使用方式）

🩺 Skin Cancer Detection (HAM10000 Dataset)

The HAM10000 dataset is a comprehensive collection of dermatoscopic images for skin lesion classification, widely used in the field of medical imaging and machine learning. It contains a diverse range of skin lesions, aimed at advancing research in dermatology, particularly for the diagnosis of skin cancers. The dataset consists of 10,000 high-resolution images of skin lesions, sourced from various individuals. This diversity helps in training robust machine learning models that can generalize well to unseen data. The primary challenge with the dataset is its significant imbalance. <br/>
Collected images are annotated and categorized into 7 classes including:

• Melanoma: 3
• Melanocytic Nevi: 1
• Basal Cell Carcinoma: 5
• Squamous Cell Carcinoma: 2
• Actinic Keratosis: 6
• Vascular Lesions: 4
• Benign Keratosis (Seborrheic Keratosis, etc.): 0
  Source: Download the HAM1000 dataset from here

💡 Class distribution

In any machine learning endeavor, it is highly recommended to perform exploratory analysis before commencing the modeling phase. This process yields insights about the data, providing valuable information that can guide and improve the modeling efforts. For example, examining the distribution of categories within the HAM10000 dataset shows that it is imbalanced, necessitating careful strategies to address this issue.
<figure style=“align: center;”>
In such cases, giving equal weight to errors across all classes can cause problems, as the model might simply classify all instances as the majority class to achieve deceptively high accuracy. This approach ignores that different cancer types need different treatments. To address this issue, it is crucial to assign error weights to each class based on their populations. For example, if the nv class is the majority, the model should prioritize it less. I use the median frequency balancing for class weighting, though mode frequency balance and inverse frequency method are also options.

🥇 Finetuning ResNet50

ResNet50, short for Residual Network with 50 layers, is a deep convolutional neural network designed to address the vanishing gradient problem in deep networks. Introduced by Kaiming He and colleagues in 2015,
[Deep Residual Learning for Image Recognition](https://arxiv

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/VahidMohammadZadehEivaghi/Skin-Cancer-Detection