Skip to main content

​
Model Architecture Ensemble

DigiPathAI employs an ensemble of three state-of-the-art deep learning architectures to achieve robust tissue segmentation in digital pathology images:

DenseNet U-Net

Dense blocks with skip connections for feature reuse

Inception-ResNet

Multi-scale feature extraction with residual connections

DeepLabv3+

Atrous spatial pyramid pooling for multi-scale context

​
Ensemble Prediction

The final segmentation is produced by combining predictions from all three models:
  1. Each model generates probability maps independently
  2. Predictions are averaged across all three architectures
  3. Optional post-processing with Conditional Random Fields (CRF) for refinement
This ensemble approach provides:
  • Robustness: Reduces model-specific biases
  • Accuracy: Leverages strengths of different architectures
  • Reliability: Consistent performance across diverse tissue types

​
Model Weights Management

DigiPathAI automatically downloads and caches pretrained model weights on first use.

​
Download Location

Model weights are stored in: 
~/.DigiPathAI/{dataset}_models/
Where {dataset} can be:
  • digestpath - DigestPath challenge dataset
  • paip - PAIP challenge dataset
  • camelyon - Camelyon challenge dataset

​
Automatic Download

The download process is handled automatically by utility functions: 
from DigiPathAI.helpers.utils import download_digestpath

# Downloads 3 model weights (~300MB total)
download_digestpath()
Downloads to: ~/.DigiPathAI/digestpath_models/
  • digestpath_deeplabv3.h5
  • digestpath_densenet.h5
  • digestpath_inception.h5

​
Loading Models

The load_trained_models() function handles model instantiation and weight loading:
helpers/utils.py (427-448)
from DigiPathAI.helpers.utils import load_trained_models

# Load individual models
inception_model = load_trained_models(
    model='inception',
    path='~/.DigiPathAI/digestpath_models/digestpath_inception.h5',
    patch_size=256
)

densenet_model = load_trained_models(
    model='dense',
    path='~/.DigiPathAI/digestpath_models/digestpath_densenet.h5',
    patch_size=256
)

deeplab_model = load_trained_models(
    model='deeplabv3',
    path='~/.DigiPathAI/digestpath_models/digestpath_deeplabv3.h5',
    patch_size=256
)
Model weights are automatically cached after first download. Subsequent runs will use cached weights without re-downloading.

​
Input Requirements

All three models expect:
  • Input shape: (height, width, 3) RGB images
  • Patch size: 256×256 pixels (default)
  • Dynamic input: Models accept (None, None, 3) for variable-sized inputs
  • Normalization: Standard ImageNet preprocessing

​
Output Format

Each model produces:
  • Shape: (height, width, 2) probability maps
  • Classes: 2-class softmax output (background, tissue)
  • Range: [0, 1] probabilities per class

​
Next Steps

DenseNet Details

Learn about DenseNet121 U-Net architecture

Inception Details

Explore Inception-ResNet-v2 U-Net

DeepLab Details

Understand DeepLabv3+ with ASPP

Build docs developers (and LLMs) love

Get started for freeTalk to us