Introduced percentile normalization for synthesis challenge metrics

GANDLF/cli/generate_metrics.py

@@ -193,6 +193,32 @@ def __fix_2d_tensor(input_tensor):
             else:
                 return input_tensor
 
+        def __percentile_clip(input_tensor, reference_tensor=None, p_min=0.5, p_max=99.5, strictlyPositive=True):
+            """Normalizes a tensor based on percentiles. Clips values below and above the percentile.
+            Percentiles for normalization can come from another tensor.
+        
+            Args:
+                input_tensor (torch.Tensor): Tensor to be normalized based on the data from the reference_tensor.
+                    If reference_tensor is None, the percentiles from this tensor will be used.
+                reference_tensor (torch.Tensor, optional): The tensor used for obtaining the percentiles.
+                p_min (float, optional): Lower end percentile. Defaults to 0.5.
+                p_max (float, optional): Upper end percentile. Defaults to 99.5.
+                strictlyPositive (bool, optional): Ensures that really all values are above 0 before normalization. Defaults to True.
+        
+            Returns:
+                torch.Tensor: The input_tensor normalized based on the percentiles of the reference tensor.
+            """
+            if(reference_tensor == None):
+                reference_tensor = input_tensor
+            v_min, v_max = np.percentile(reference_tensor, [p_min,p_max]) #get p_min percentile and p_max percentile
+
+            if( v_min < 0 and strictlyPositive): #set lower bound to be 0 if it would be below
+                v_min = 0
+            output_tensor = np.clip(input_tensor,v_min,v_max) #clip values to percentiles from reference_tensor
+            output_tensor = (output_tensor - v_min)/(v_max-v_min) #normalizes values to [0;1]
+
+            return output_tensor
+
         for _, row in tqdm(input_df.iterrows(), total=input_df.shape[0]):
             current_subject_id = row[headers["subjectid"]]
             overall_stats_dict[current_subject_id] = {}
@@ -219,9 +245,9 @@ def __fix_2d_tensor(input_tensor):
             # Normalize to [0;1] based on GT (otherwise MSE will depend on the image intensity range)
             normalize = parameters.get("normalize", True)
             if normalize:
-                v_max = gt_image_infill.max()
-                output_infill /= v_max
-                gt_image_infill /= v_max
+                reference_tensor = gt_image * ~mask #use all the tissue that is not masked for normalization
+                gt_image_infill = __percentile_clip(gt_image_infill, reference_tensor=reference_tensor, p_min=0.5, p_max=99.5, strictlyPositive=True)
+                prediction_infill = __percentile_clip(prediction_infill, reference_tensor=reference_tensor, p_min=0.5, p_max=99.5, strictlyPositive=True)
 
             overall_stats_dict[current_subject_id][
                 "ssim"
@@ -258,6 +284,7 @@ def __fix_2d_tensor(input_tensor):
                 gt_image_infill, output_infill
             ).item()
 
+            #TODO: use data_range=1.0 as parameter for PSNR when the Pull request is accepted that introduces the data_range parameter!
             # PSNR - similar to pytorch PeakSignalNoiseRatio until 4 digits after decimal point
             overall_stats_dict[current_subject_id]["psnr"] = peak_signal_noise_ratio(
                 gt_image_infill, output_infill