Histopathologic Cancer Detection using Attention Mechanism

Deep learning model for binary classification of histopathologic images to detect metastatic cancer in lymph node sections using ResNet18 with attention mechanism

Dataset

Source: HCD-Cropped Dataset on Kaggle

Dataset Characteristics

• Total Images: 220,025 histopathologic images
• Format: TIF images (32×32×3 pixels)
• Classes: Binary classification
  • Class 0: No metastatic tissue
  • Class 1: Metastatic tissue present
• Image Resolution: 32×32 pixels (upscaled to 224×224 for ResNet18 input)
• Channels: 3 (RGB)
• Split: 80% training, 10% validation, 10% test

The dataset contains cropped patches from whole-slide images of lymph node sections, where the task is to identify metastatic tissue

Model Architecture

Attention-Enhanced ResNet18

The model uses a ResNet18 backbone pretrained on ImageNet with a custom attention module for interpretable feature learning

Architecture Details

• Backbone: ResNet18 (pretrained on ImageNet1K)
• Feature Extractor: All ResNet18 layers except final FC layer
• Feature Dimension: 512 → 16 (compressed representation)
• Attention Mechanism: Tanh-based attention with softmax normalization
• Dropout: 0.5 for regularization
• Output: 2 classes (binary classification)

class AttentionModule(nn.Module):
    def __init__(self):
        super(AttentionModule, self).__init__()
        original_model = models.resnet18(weights=ResNet18_Weights.IMAGENET1K_V1)
        self.features = nn.Sequential(*list(original_model.children())[:-1])
        self.fc1 = nn.Linear(512, 16)
        self.dropout = nn.Dropout(0.5)
        self.attention = nn.Sequential(
            nn.Linear(16, 16),
            nn.Tanh(),
            nn.Linear(16, 1),
            nn.Softmax(dim=1)
        )
        self.classifier = nn.Linear(16, 2)

Total Parameters: 11,185,043 (all trainable)

Data Preprocessing

• Resize: 32×32 → 224×224 pixels
• Normalization: ImageNet statistics (mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
• Tensor conversion with automatic augmentation support

Training Configuration

• Optimizer: Adam (lr=0.001)
• Loss Function: CrossEntropyLoss
• Batch Size: 32
• Epochs: 6
• Device: CUDA (GPU acceleration)

Performance

Training Progress

Epoch	Train Loss	Train Acc	Val Loss	Val Acc
1	0.4271	81.03%	0.4236	81.07%
2	0.3761	83.84%	0.3545	84.75%
3	0.3456	85.43%	0.3393	85.24%
4	0.3235	86.48%	0.3150	86.45%
5	0.3051	87.35%	0.3158	86.66%
6	0.2866	88.21%	0.3242	86.25%

Final Test Performance

• Test Loss: 0.3130
• Test Accuracy: 86.71%

Installation

# Clone the repository
git clone https://github.com/yourusername/histopathologic-cancer-detection.git
cd histopathologic-cancer-detection

# Install dependencies
pip install torch torchvision torchinfo pandas pillow rasterio scikit-learn

Usage

Training

# Load dataset
cancer_dataset = CancerDataset(image_dir=image_dir, labels_df=labels_df, transform=transform)

# Split data
train_dataset, val_dataset, test_dataset = torch.utils.data.random_split(
    cancer_dataset, [0.8, 0.1, 0.1]
)

# Initialize model
model = AttentionModule().to(device)
optimizer = optim.Adam(model.parameters(), lr=0.001)
criterion = nn.CrossEntropyLoss()

# Train
for epoch in range(num_epochs):
    train_loss, train_acc = train_model(model, train_loader, device, optimizer, criterion)
    val_loss, val_acc = validate_model(model, val_loader, device, criterion)

Inference

# Load trained weights
model.load_state_dict(torch.load('6-epoch-cancer_detection_weights.pth'))
model.eval()

# Get predictions with attention weights
outputs, attention_weights = model(images, return_attention=True)

Key Features

• ✅ Attention mechanism for interpretable predictions
• ✅ Transfer learning with ImageNet-pretrained ResNet18
• ✅ Efficient training on 220K+ images
• ✅ 86.71% test accuracy
• ✅ Attention weight visualization support

Requirements

torch>=2.0.0
torchvision>=0.15.0
torchinfo
pandas
pillow
rasterio
scikit-learn

Acknowledgments

• Dataset: HCD-Cropped on Kaggle
• Base Architecture: ResNet18 from torchvision.models