Merge branch 'master' of github.com:jmlipman/MedicDeepLabv3Plus

jmlipman · Apr 15, 2021 · 27b5dcc · 27b5dcc
2 parents 6c84796 + f2e6528
commit 27b5dcc
Show file tree

Hide file tree

Showing 2 changed files with 27 additions and 25 deletions.
diff --git a/README.md b/README.md
@@ -1,6 +1,10 @@
 MedicDeepLabv3+
 ======================
 
+[![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.4009212.svg)](https://doi.org/10.5281/zenodo.4009212) (Software)
+
+[![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.4009246.svg)](https://doi.org/10.5281/zenodo.4009246) (Trained models)
+
 Repository of MedicDeepLabv3+
 
 ![Architecture](medicdeeplabv3plus.png "MedicDeepLabv3+ Architecture")
@@ -30,23 +34,19 @@ This implementation of MedicDeepLabv3+ allows combining several models trained s
 .
  ├── eval.py # Generates segmentation masks. It requires a file with the trained parameters of MedicDeepLabv3+ (provided by train.py)
  ├── train.py # Optimizes MedicDeepLabv3+ and saves its optimized parameters (required by eval.py)
- ├── lib
- │   ├── losses.py # Cross Entropy + Dice Loss functions
- │   ├── metric.py # Metrics to quantify segmentations quality i.e. Dice coeff., Hausdorff distance, Compactness
- │   ├── utils.py # Other functions.
- │   ├── blocks
- │   │   ├── BasicBlocks.py # Contains basic operations of the ConvNet
- │   │   └── MedicDeepLabv3PlusBlocks.py # Blocks of operations for MedicDeepLabv3+
- │   ├── data
- │   │   ├── BaseDataset.py # Basic dataset operations
- │   │   └── DataWrapper.py # Reads and parses the NIfTI files
- │   └── models
- │      ├── BaseModel.py # Contains main training and evaluation procedures
- │      └── MedicDeepLabv3Plus.py # Pytorch definition of our model
- └── trained_models # Trained MedicDeepLabv3+ parameters
-     ├── MedicDeepLabv3Plus-model-300_1
-     ├── MedicDeepLabv3Plus-model-300_2
-     └── MedicDeepLabv3Plus-model-300_3
+ └── lib
+     ├── losses.py # Cross Entropy + Dice Loss functions
+     ├── metric.py # Metrics to quantify segmentations quality i.e. Dice coeff., Hausdorff distance, Compactness
+     ├── utils.py # Other functions.
+     ├── blocks
+     │   ├── BasicBlocks.py # Contains basic operations of the ConvNet
+     │   └── MedicDeepLabv3PlusBlocks.py # Blocks of operations for MedicDeepLabv3+
+     ├── data
+     │   ├── BaseDataset.py # Basic dataset operations
+     │   └── DataWrapper.py # Reads and parses the NIfTI files
+     └── models
+        ├── BaseModel.py # Contains main training and evaluation procedures
+        └── MedicDeepLabv3Plus.py # Pytorch definition of our model
 ```
 
 ### 2. Installation and Requirements
@@ -63,7 +63,7 @@ This implementation of MedicDeepLabv3+ allows combining several models trained s
 
 1. Install dependencies with pip
 ```cshell
-pip install scipy, scikit-image, nibabel
+pip install scipy scikit-image nibabel
 ```
 
 2. Download source code
@@ -154,6 +154,8 @@ Optionally, if the ground truth is provided in the same folder as the data (as d
 
 Finally, you can choose in which GPU to execute eval.py, as in train.py. Since eval.py does not take too long to execute, unlike train.py, evaluating MedicDeepLabv3+ on the CPU with --gpu -1 is a viable option.
 
+Note: we provide three trained MedicDeepLabv3+ models at DOI:10.5281/zenodo.4009246
+
 ```cshell
 # Minimum training setup
 python eval.py --input DIR --output DIR --model FILE(S)
@@ -182,4 +184,4 @@ Files generated by eval.py:
 [MIT License](LICENSE)
 
 ### 5. Contact
-Feel free to write an email with questions or feedback about MedicDeepLabv3+ at **juanmiguel.valverde@uef.com**
+Feel free to write an email with questions or feedback about MedicDeepLabv3+ at **juanmiguel.valverde@uef.fi**
diff --git a/lib/models/BaseModel.py b/lib/models/BaseModel.py
@@ -176,13 +176,13 @@ def evaluate(self, test_loader, metrics, remove_islands, save_output=True):
                 if remove_islands:
                     y_pred_cpu = removeSmallIslands(y_pred_cpu, thr=20)
 
-                # If GT was provided
+                # Predictions (and GT) separate the two hemispheres
+                # combineLabels will combine these such that it creates
+                # brainmask and contra-hemisphere ROIs instead of
+                # two different hemisphere ROIs.
+                y_pred_cpu = combineLabels(y_pred_cpu)
+                # If GT was provided it measures the performance
                 if len(y_true_cpu.shape) > 1:
-                    # Predictions (and GT) separate the two hemispheres
-                    # combineLabels will combine these such that it creates
-                    # brainmask and contra-hemisphere ROIs instead of
-                    # two different hemisphere ROIs.
-                    y_pred_cpu = combineLabels(y_pred_cpu)
                     y_true_cpu = combineLabels(y_true_cpu)
 
                     results[id_] = Measure.all(y_pred_cpu, y_true_cpu)