Model Weights for the ISGPP model for the segmentation of Pancreatic Ductal Adenocarcinoma (PDAC) after neoadjuvant therapy
This repository contains the model weights for the ISGPP segmentation model, which segments cancer tissue in whole slide images of H&E sections from PDAC samples after neoadjuvant therapy. This model is described in the following publication: https://journals.lww.com/ajsp/fulltext/9900/artificial_intelligence_based_segmentation_of.380.aspx
Please cite this paper if you are using the ISGPP model for your research.
The model is based on a U-net with a densenet161 encoder. The model was defined in the segmentation-models for PyTorch library (https://pypi.org/project/segmentation-models-pytorch/).
For training, 528 digitized histopathological whole slide images (hematoxylin & eosin staining) from resected pancreatic cancer specimen from 14 centers in 7 countries in Europe, North America, Australia, and Asia were included. Four different scanner types were used: Philips, Hamamatsu, 3DHistech and Leica. More details about the dataset and model training can be found in our publication: https://journals.lww.com/ajsp/fulltext/9900/artificial_intelligence_based_segmentation_of.380.aspx
The model was trained using binary-cross-entropy-loss with ADAM optimizer for 40 epochs. The learning rate was set to 0.0001, and a learning rate decay every 5 epochs.
The meta-analysis of the cross-validations showed a mean F1 score of 0.78 (0.71 – 0.84).
Size inputpatches: 512 * 512 * 3
Resolution inputpatches: 2px/micron
Normalize inputpatches using: torch.transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
Example code for loading the model:
import torch
import segmentation_models_pytorch as smp
# Model definition
aux_params = dict(pooling='max', classes=5)
model = smp.Unet("densenet161", classes=5, aux_params=aux_params)
# Load model weights
model.load_state_dict(torch.load("checkpoint.pt"))
model.eval()
# example inference
pred, _ = model(image_loader(patch))
Output:
Class 0: background
class 1: Normal ducts
Class 2: Cancer ducts
Class 3: residual parenchyma
Class 4: fat
For more information:
Boris Janssen (b.v.janssen@amsterdamumc.nl) or Onno de Boer (o.j.deboer@amsterdamumc.nl)