From c8ce759e9924d992e96faf3cac7e0aef68ea4040 Mon Sep 17 00:00:00 2001
From: Ajey Pai K <ajeypie@gmail.com>
Date: Mon, 5 Aug 2024 11:50:13 +0200
Subject: [PATCH] Add documentation for training tissue foreground

---
 docs/model_zoo.rst | 58 ++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 58 insertions(+)

diff --git a/docs/model_zoo.rst b/docs/model_zoo.rst
index b62455a..4533b27 100644
--- a/docs/model_zoo.rst
+++ b/docs/model_zoo.rst
@@ -36,3 +36,61 @@ Segmentation models
      - H&E
      - `download <https://files.aiforoncology.nl/ahcore/v0.1.1/tissue_background.zip>`_
      - v0.1.1
+
+Training tissue foreground segmentation model
+=============================================
+
+Introduction
+------------
+This file documents the training of a tissue foreground segmentation model. The model is trained on a dataset consisting of whole slide images (WSIs) of Camelyon16 and TCGA-BRCA datasets.
+
+The following steps are necessary to train the model:
+1. Creation of a manifest file containing the paths to the WSIs and their corresponding annotations.
+2. Configuration of ahcore to train the foreground segmentation model.
+
+Create a manifest file
+----------------------
+In order to create a manifest file to train a segmentation model, use the cli utility provided here `<https://github.com/NKI-AI/pathology-data-utilities/tree/main/ahcore/tissue_background>`_.
+
+Configure ahcore for training the model
+---------------------------------------
+Use the following data_description to train the model:
+
+```
+_target_: ahcore.utils.data.DataDescription
+data_dir: ${oc.env:DATA_DIR}
+annotations_dir: ${oc.env:ANNOTATIONS_DIR} # specify in .env
+manifest_database_uri: sqlite:////${oc.env:MANIFEST_PATH}/tcga_database.db
+mask_label: null
+mask_threshold: null
+manifest_name: "v20231115_tcga_camelyon_bg"
+split_version: "v1"
+roi_name: null  # This is the name of the label that carries the ROI
+use_roi: False
+apply_color_profile: False
+# Eg. 512 x 512 is: tile_size 726 x 726 and tile_overlap 107 x 107
+# Tiles are cropped in the transforms downstream, to ensure the patch is completely visible, they are extracted
+# slightly larger qrt 2) with sufficient overlap so we have a 512 stride.
+training_grid:
+  mpp: 8
+  output_tile_size: [726, 726]
+  tile_overlap: [0, 0]
+  tile_size: [512, 512]
+inference_grid:
+  mpp: 8
+  tile_size: [512, 512]
+  tile_overlap: [128, 128]
+
+num_classes: 2
+use_class_weights: false  # Use the class weights in the loss
+
+remap_labels:
+  specimen: specimen
+
+index_map:
+  specimen: 1
+
+color_map:
+  0: "yellow"
+  1: "red"
+```