Skip to main content

​
Overview

DigiPathAI provides specialized models for colon cancer segmentation, trained on the DigestPath dataset. These models detect and segment cancerous regions in colorectal histopathology slides.

​
DigestPath Dataset

The DigestPath dataset contains annotated whole slide images of colorectal cancer tissue. The model ensemble includes:

Inception

Multi-scale feature extraction for diverse tissue patterns

DenseNet

Dense connections for efficient feature learning

DeepLabV3

Precise boundary segmentation with atrous convolution

​
Model Details

The colon cancer models are automatically managed by DigiPathAI: 
# From Segmentation.py lines 232-246
if mode == 'colon':
    path = os.path.join(home, '.DigiPathAI/digestpath_models')
    if (not os.path.exists(os.path.join(path, 'digestpath_inception.h5'))) or \
        (not os.path.exists(os.path.join(path, 'digestpath_deeplabv3.h5'))) or \
        (not os.path.exists(os.path.join(path, 'digestpath_densenet.h5'))):
        if status is not None: status['status'] = "Downloading Trained Models"
        download_digestpath() 
        model_path_inception = os.path.join(path, 'digestpath_inception.h5')
        model_path_deeplabv3 = os.path.join(path, 'digestpath_deeplabv3.h5')
        model_path_densenet = os.path.join(path, 'digestpath_densenet.h5')
    else :
        if status is not None: status['status'] = "Found Trained Models, Skipping download"
        model_path_inception = os.path.join(path, 'digestpath_inception.h5')
        model_path_deeplabv3 = os.path.join(path, 'digestpath_deeplabv3.h5')
        model_path_densenet = os.path.join(path, 'digestpath_densenet.h5')
Models are stored in ~/.DigiPathAI/digestpath_models/ and include:
  • digestpath_inception.h5
  • digestpath_deeplabv3.h5
  • digestpath_densenet.h5

​
Usage Example

​
Basic Colon Segmentation

from DigiPathAI.Segmentation import getSegmentation

# Segment colon cancer tissue
prediction = getSegmentation(
    img_path='colon_slide.tiff',
    mode='colon',
    patch_size=256,
    stride_size=128,
    batch_size=32,
    quick=False  # Use ensemble of all three models
)

​
Quick Mode (Single Model)

# Use only DenseNet for faster inference
prediction = getSegmentation(
    img_path='colon_slide.tiff',
    mode='colon',
    model='dense',
    quick=True
)

​
With Test-Time Augmentation

# Improve accuracy with TTA
prediction = getSegmentation(
    img_path='colon_slide.tiff',
    mode='colon',
    tta_list=['FLIP_LEFT_RIGHT', 'ROTATE_90'],
    quick=False
)
For production use with colon tissue, set quick=False to leverage the full ensemble. This provides better accuracy and uncertainty estimates.

​
Parameters

ParameterDescriptionDefaultColon-Specific Notes
modeTissue type-Must be set to 'colon'
modelArchitecture choice'dense'Options: ‘dense’, ‘inception’, ‘deeplabv3’
quickSingle vs ensembleTrueSet to False for ensemble
patch_sizeInference patch size256Recommended: 256 for colon
stride_sizeSliding window stride128Smaller values = more overlap
batch_sizeBatch size for inference32Adjust based on GPU memory

​
Output

The segmentation returns three files:
  1. Probability Map - Continuous probability values (0-1)
  2. Binary Mask - Thresholded segmentation (threshold=0.3)
  3. Uncertainty Map - Model variance for quality assessment
# Outputs are saved as pyramidal TIFF files
# - probs_path: Probability heatmap
# - mask_path: Binary segmentation mask  
# - uncertainty_path: Prediction uncertainty

Build docs developers (and LLMs) love

Get started for freeTalk to us