Add dice and jaccard to postprocessing

Author: youennmereljourdan

README.md

@@ -11,21 +11,24 @@ See https://github.com/Inria-Empenn/fmri_feature_model/tree/master
 
 # Describe your data
 
-
+Modify `data_desc_sample.json` to your usage
 
 # Run configurations
 
 - Run configurations list from `/config.csv` file.
 - Read data from `/data`
 - Write NiPype cache in `/workdir`
-- Write results in `/results`
+- Write results to `/results`
 
 ## Locally
 Change `/local/path/to/...` to your local paths
 ``` sh
 docker run -u root -v "/local/path/to/data:/data" -v "/local/path/to/results:/results" -v "/local/path/to/workdir:/work" -v "/local/path/to/configs:/configs" fmri-conf-runner python -u run.py --configs "/configs/config.csv" --data /data/data_desc.json --ref /configs/config_ref.csv
 ```
 ## On Abaca (Inria cluster)
+
+Use `run_configs.sh`
+
 - Start 40 jobs in parallel.
 - Timeout for each job is 4h
 - Needs 40 configurations named CSV `config_[1..40].csv`
@@ -34,9 +37,19 @@ oarsub -S -n fmri-conf-runner ./run_configs.sh
 ```
 
 # Postprocess data
+- Read data from `/results`
+- Write results to `/results`
+  - 'mean_result.nii' : mean statistic map
+  - 'correlations.csv' : pairwise correlations
+  - 'dataset.csv' : 
 ## Locally
 Change `/local/path/to/...` to your local paths
 ``` sh
 docker run -u root -v "/local/path/to/results:/results" fmri-conf-runner python -u postprocess.py --results "/results"
 ```
 ## On Abaca (Inria cluster)
+
+On Abaca (Inria cluster), use `postprocess.sh`
+``` sh
+oarsub -S -n postprocess ./postprocess.sh
+```

data_desc_hcp_motor.json

@@ -0,0 +1,15 @@
+{
+  "data_path" : "/home/user/fmri-feature-model/data/auditory",
+  "input" : {
+    "anat" : "sub-{subject_id}/anat/sub-{subject_id}_T1w.nii",
+    "func" : "sub-{subject_id}/func/sub-{subject_id}_task-auditory_bold.nii",
+    "events" : "sub-{subject_id}/func/sub-{subject_id}_task-auditory_events_corrected.tsv"
+  },
+  "result_path" : "/home/user/fmri-conf-runner/results",
+  "work_path" : "/home/user/fmri-conf-runner/work",
+  "task" : "auditory",
+  "subjects" : ["01"],
+  "slices_nb" : 64,
+  "tr" : 7.0,
+  "units" : "scans"
+}

data_desc_auditory.json renamed to data_desc_sample.json

@@ -56,9 +56,55 @@ def get_correlation_coefficient(self,
             return corrcoef(data_1, data_2)[0][1]
         if method == 'spearman':
             return spearmanr(data_1, data_2).correlation
+        if method == 'dice':
+            return self.get_dice(data_1, data_2)
+        if method == 'jaccard':
+            return self.get_jaccard(data_1, data_2)
 
         raise AttributeError(f'Wrong correlation method provided: {method}.')
 
+    def get_dice(self, data_1: np.ndarray, data_2: np.ndarray) -> float:
+        """
+        Return the DICE score between two images data
+        :param data_1:
+        :param data_2:
+        :return:
+        """
+        # Binarize
+        binary1 = data_1 > 0
+        binary2 = data_2 > 0
+
+        # Compute intersection + volumes
+        intersection = np.logical_and(binary1, binary2).sum()
+        volume = binary1.sum() + binary2.sum()
+
+        # Avoid divide by 0
+        if volume == 0:
+            return 0.0
+
+        return (2. * intersection) / volume
+
+    def get_jaccard(self, data_1: np.ndarray, data_2: np.ndarray) -> float:
+        """
+        Return the Jaccard Index between two images data
+        :param data_1:
+        :param data_2:
+        :return:
+        """
+        # Binarize
+        binary1 = data_1 > 0
+        binary2 = data_2 > 0
+
+        # Compute intersection + union
+        intersection = np.logical_and(binary1, binary2).sum()
+        union = np.logical_or(binary1, binary2).sum()
+
+        # Avoid division by zero
+        if union == 0:
+            return 0.0
+
+        return intersection / union
+
     def mask_using_zeros(self, data_image: Nifti1Image) -> Nifti1Image:
         """ Mask an image by replacing NaNs with zeros.
 

postprocess/correlation_service.py

@@ -40,20 +40,26 @@ def get_all_correlations(self, path, ids: List[str]) -> pd.DataFrame:
             for j in range(i, n):
                 if i == j:
                     corr = 1.0
+                    dice = 1.0
+                    jacc = 1.0
                 else:
                     tgt = os.path.join(path, ids[j], '_subject_id_01', 'result.nii')
                     corr = self.corr_srv.get_correlation_coefficient(src, tgt, 'spearman')
-                data.append((ids[i], ids[j], corr))
+                    dice = self.corr_srv.get_correlation_coefficient(src, tgt, 'dice')
+                    jacc = self.corr_srv.get_correlation_coefficient(src, tgt, 'jaccard')
+                data.append((ids[i], ids[j], corr, dice))
                 if i != j:
-                    data.append((ids[j], ids[i], corr))
+                    data.append((ids[j], ids[i], corr, dice, jacc))
             # Add correlation from mean
             mean = os.path.join(path, 'mean_result.nii')
             corr = self.corr_srv.get_correlation_coefficient(src, mean, 'spearman')
-            data.append((ids[i], 'mean', corr))
-            data.append(('mean', ids[i], corr))
+            dice = self.corr_srv.get_correlation_coefficient(src, tgt, 'dice')
+            jacc = self.corr_srv.get_correlation_coefficient(src, tgt, 'jaccard')
+            data.append((ids[i], 'mean', corr, dice, jacc))
+            data.append(('mean', ids[i], corr, dice, jacc))
             print(f"Processed correlations for [{i+1} / {n}] result")
-        data.append(('mean', 'mean', 1.0))
-        dataframe = pd.DataFrame(data, columns=['source', 'target', 'correlation'])
+        data.append(('mean', 'mean', 1.0, 1.0, 1.0))
+        dataframe = pd.DataFrame(data, columns=['source', 'target', 'spearman', 'dice', 'jaccard'])
         return dataframe.sort_values(by='correlation', ascending=False)
 
     def get_mean_image(self, inputs: list, batch_size: int) -> nib.Nifti1Image: