Merge pull request #237 from dzenanz/master

Remove SNR and CNR from training and inference
OpenImaging · Nov 29, 2021 · dd222c0 · dd222c0
2 parents 8bb7f8f + 95fe0a5
commit dd222c0
Show file tree

Hide file tree

Showing 8 changed files with 41 additions and 27 deletions.
diff --git a/miqa/learning/correlator.py b/miqa/learning/correlator.py
@@ -0,0 +1,26 @@
+#!/usr/bin/env python3
+
+import pandas as pd
+from sklearn.metrics import confusion_matrix
+
+df = pd.read_csv('M:/MIQA/data.csv')  # manually converted TRUE/FALSE into 1/0
+print(f'count NaN: {df.isnull().sum().sum()}')
+correlation_df = df.corr()
+correlation_df.to_csv('M:/MIQA/correlations2.csv')
+print(correlation_df)
+
+cm = pd.DataFrame(confusion_matrix(df['overall_qa_assessment'], df['cnr']))
+cm.to_csv('M:/MIQA/oQA_CNR.csv')
+print(cm)
+
+cm = pd.DataFrame(confusion_matrix(df['overall_qa_assessment'], df['snr']))
+cm.to_csv('M:/MIQA/oQA_SNR.csv')
+print(cm)
+
+# df = pd.read_csv('M:/MIQA/PredictHD_small/phenotype/bids_image_qc_information.tsv', sep='\t')
+# df = df.drop(columns=['participant_id', 'session_id', 'series_number'])
+# df.to_csv('M:/MIQA/dataQA.csv')
+# print(f"count NaN: {df.isnull().sum().sum()}")
+# correlation_df = df.corr()
+# correlation_df.to_csv('M:/MIQA/correlations.csv')
+# print(correlation_df)
diff --git a/miqa/learning/models/miqaT1-val0.pth b/miqa/learning/models/miqaT1-val0.pth
diff --git a/miqa/learning/models/miqaT1-val1.pth b/miqa/learning/models/miqaT1-val1.pth
diff --git a/miqa/learning/models/miqaT1-val2.pth b/miqa/learning/models/miqaT1-val2.pth
diff --git a/miqa/learning/models/miqaT1-val3.pth b/miqa/learning/models/miqaT1-val3.pth
diff --git a/miqa/learning/models/miqaT1-val4.pth b/miqa/learning/models/miqaT1-val4.pth
diff --git a/miqa/learning/nn_inference.py b/miqa/learning/nn_inference.py
@@ -10,7 +10,7 @@
 
 logger = logging.getLogger(__name__)
 
-regression_count = 3  # first 3 values are overall QA, SNR and CNR
+regression_count = 1  # use QA, ignore SNR and CNR
 artifacts = [
     'normal_variants',
     'lesions',
@@ -154,11 +154,7 @@ def evaluate_model(model, data_loader, device, writer, epoch, run_name):
 
 
 def label_results(result):
-    labeled_results = {
-        'overall_quality': clamp(result[0] / 10.0, 0.0, 1.0),
-        'signal_to_noise_ratio': clamp(result[1] / 10.0, 0.0, 1.0),
-        'contrast_to_noise_ratio': clamp(result[2] / 10.0, 0.0, 1.0),
-    }
+    labeled_results = {'overall_quality': clamp(result[0] / 10.0, 0.0, 1.0)}
     for artifact_name, value in zip(artifacts, result[regression_count:]):
         labeled_results[artifact_name] = clamp(value, 0.0, 1.0)
     return labeled_results

diff --git a/miqa/learning/nn_training.py b/miqa/learning/nn_training.py
@@ -163,16 +163,8 @@ def forward(self, output, target):
         qa_target = target[..., 0]
         qa_loss = torch.mean((qa_out - qa_target) ** 2)
 
-        snr_out = output[..., 1]
-        snr_target = target[..., 1]
-        snr_loss = torch.mean((snr_out - snr_target) ** 2)
-
-        cnr_out = output[..., 2]
-        cnr_target = target[..., 2]
-        cnr_loss = torch.mean((cnr_out - cnr_target) ** 2)
-
-        # overall QA is more important than SNR and CNR
-        loss = 10 * qa_loss + snr_loss + cnr_loss
+        # overall QA is more important than individual artifacts
+        loss = 10 * qa_loss
 
         for i in range(self.presence_count):
             i_target = target[..., i + regression_count]
@@ -213,7 +205,7 @@ def create_train_and_test_data_loaders(df, count_train):
         if exists:
             images.append(row.file_path)
 
-            row_targets = [row.overall_qa_assessment, row.snr, row.cnr]
+            row_targets = [row.overall_qa_assessment]
             for i in range(len(artifacts)):
                 artifact_value = row[artifact_column_indices[i]]
                 converted_result = convert_bool_to_int(artifact_value)