Upload 2 files

Util/Custom_Model.py

@@ -0,0 +1,44 @@
+import timm
+import torch.nn as nn
+
+class Build_Custom_Model(nn.Module):
+    def __init__(self, model_name, target_size, pretrained=False):
+        super().__init__()
+        self.model = timm.create_model(model_name, pretrained=pretrained, in_chans=1)
+        if(model_name=="vit_base_patch16_224" or model_name=="swin_base_patch4_window7_224"):
+            self.n_features = self.model.head.in_features
+            self.model.head = nn.Linear(self.n_features, target_size)
+        if(model_name=="resnet34d"):
+            self.n_features = self.model.fc.in_features
+            self.model.fc = nn.Linear(self.n_features, target_size)
+        if(model_name=="resnet18d"):
+            self.n_features = self.model.fc.in_features
+            self.model.fc = nn.Linear(self.n_features, target_size)
+        if(model_name=="tf_efficientnet_b7_ns"):
+            self.n_features = self.model.classifier.in_features
+            self.model.classifier = nn.Linear(self.n_features, target_size)
+        if(model_name=="tf_efficientnet_b0_ns"):
+            self.n_features = self.model.classifier.in_features
+            self.model.classifier = nn.Linear(self.n_features, target_size)
+        if(model_name=="tf_efficientnet_lite0"):
+            self.n_features = self.model.classifier.in_features
+            self.model.classifier = nn.Linear(self.n_features, target_size)
+        if(model_name=="mobilenetv2_050"):
+            self.n_features = self.model.classifier.in_features
+            self.model.classifier = nn.Linear(self.n_features, target_size)
+        if(model_name=="eca_nfnet_l0"):
+            self.n_features = self.model.head.fc.in_features
+            self.model.head.fc = nn.Linear(self.n_features, target_size)
+
+    def forward(self, x):
+        output = self.model(x)
+        return output
+
+def reshape_transform(tensor, height=7, width=7):
+    result = tensor.reshape(tensor.size(0),
+        height, width, tensor.size(2))
+
+    # Bring the channels to the first dimension,
+    # like in CNNs.
+    result = result.transpose(2, 3).transpose(1, 2)
+    return result

Util/DICOM.py

@@ -0,0 +1,81 @@
+import cv2
+import numpy as np
+
+class DICOM_Utils(object):
+    def apply_windowing(image_array, window_center, window_width):
+        """
+        Apply windowing to a DICOM image array.
+
+        Parameters:
+        - image_array: numpy array of the DICOM image
+        - window_center: center of the window
+        - window_width: width of the window
+
+        Returns:
+        - Windowed image array
+        """
+        lower_bound = window_center - (window_width / 2)
+        upper_bound = window_center + (window_width / 2)
+
+        # Apply windowing
+        windowed_image = image_array.copy()
+        windowed_image[windowed_image < lower_bound] = lower_bound
+        windowed_image[windowed_image > upper_bound] = upper_bound
+
+        # Normalize to [0, 255]
+        windowed_image = ((windowed_image - lower_bound) / window_width) * 255
+
+        return windowed_image.astype('uint8')
+
+    def transform_image_for_display(image_array):
+        """
+        Transform the image for display: Flip horizontally and then rotate 90 degrees to the right.
+
+        Parameters:
+        - image_array: numpy array of the image
+
+        Returns:
+        - Transformed image array
+        """
+        # Flip horizontally
+        flipped_image = np.fliplr(image_array)
+
+        # Rotate 90 degrees to the right
+        rotated_image = np.rot90(flipped_image, 1)
+
+        return rotated_image
+
+    def apply_CAM_overlay(heatmap, windowed_image, overlay_alpha=0.4):
+        """
+        Apply CAM (Class Activation Map) overlay to a given image.
+
+        Parameters:
+        - heatmap: torch.Tensor, the heatmap generated by CAM.
+        - windowed_image: numpy.ndarray, the windowed image to overlay the heatmap on.
+        - overlay_alpha: float, the transparency for overlaying heatmap. Default is 0.4.
+
+        Returns:
+        - overlayed: numpy.ndarray, the resulting image after overlaying the heatmap.
+        """
+        # Convert the heatmap tensor to a numpy array
+        heatmap_np = heatmap.cpu().numpy().squeeze()
+
+        # Normalize the heatmap to [0, 255]
+        heatmap_normalized = ((heatmap_np - heatmap_np.min()) / 
+                            (heatmap_np.max() - heatmap_np.min()) * 255).astype(np.uint8)
+
+        # Convert the normalized heatmap to a colormap (for example, using the "jet" colormap)
+        heatmap_colormap = cv2.applyColorMap(heatmap_normalized, cv2.COLORMAP_JET)
+
+        # Resize the colormap to the original image size
+        heatmap_resized = cv2.resize(heatmap_colormap, 
+                                    (windowed_image.shape[1], windowed_image.shape[0]))
+
+        # Convert the grayscale windowed_image to 3 channels
+        windowed_image_colored = cv2.cvtColor(windowed_image, cv2.COLOR_GRAY2BGR)
+
+        # Overlay the heatmap on the original image with a certain transparency
+        overlayed = cv2.addWeighted(windowed_image_colored, 1 - overlay_alpha, 
+                                    heatmap_resized, overlay_alpha, 0)
+
+        return overlayed