Updated Model Training

Datasets/Sample_Dataset/val/Bladder1.json

@@ -0,0 +1,10 @@
+{
+  "num_total": 183,
+  "num_pos": 15,
+  "num_neg": 168,
+  "percent_pos": 8.2,
+  "prob_thresh": 150,
+  "size_thresh": 30,
+  "size_thresh_upper": null,
+  "marker_thresh": null
+}

Datasets/Sample_Dataset/val/Lung1.json

@@ -0,0 +1,10 @@
+{
+  "num_total": 60,
+  "num_pos": 11,
+  "num_neg": 49,
+  "percent_pos": 18.3,
+  "prob_thresh": 150,
+  "size_thresh": 78,
+  "size_thresh_upper": null,
+  "marker_thresh": null
+}

Datasets/Sample_Dataset/val/metrics.json

@@ -0,0 +1,22 @@
+{
+  "Lung1": {
+    "num_total": 60,
+    "num_pos": 11,
+    "num_neg": 49,
+    "percent_pos": 18.3,
+    "prob_thresh": 150,
+    "size_thresh": 78,
+    "size_thresh_upper": null,
+    "marker_thresh": null
+  },
+  "Bladder1": {
+    "num_total": 183,
+    "num_pos": 15,
+    "num_neg": 168,
+    "percent_pos": 8.2,
+    "prob_thresh": 150,
+    "size_thresh": 30,
+    "size_thresh_upper": null,
+    "marker_thresh": null
+  }
+}

Image_Processing/Augmentation.py

@@ -6,12 +6,13 @@
 
 
 class Augmentation:
-    def __init__(self, images):
+    def __init__(self, images, tile_size=512):
         self.images = images
         self.shape = self.images[list(self.images.keys())[0]].shape
         self.rotation_angle = np.random.choice([0, 90, 180, 270], 1)[0]
         # self.zoom_value = random.randint(0, 5)
         self.alpha_affine = 0.1
+        self.tile_size = tile_size
 
     def pipeline(self):
         """
@@ -30,12 +31,13 @@ def zoom(self):
         :return:
         """
         new_size = random.randint(int(self.shape[0] * 0.75), self.shape[0])
+        assert self.shape[1] - new_size >= 0, f'self.shape[1] - new_size ({self.shape[1]} - {new_size})should not be negative'
         start_point = (random.randint(0, self.shape[0] - new_size), random.randint(0, self.shape[1] - new_size))
         for key in self.images.keys():
             try:
-                self.images[key] = cv2.resize(self.images[key][start_point[0]: start_point[0] + new_size, start_point[1]: start_point[1] + new_size], (512, 512))
-            except:
-                print(key + ' not available')
+                self.images[key] = cv2.resize(self.images[key][start_point[0]: start_point[0] + new_size, start_point[1]: start_point[1] + new_size], (self.tile_size, self.tile_size))
+            except Exception as e:
+                print(e)
 
     def rotate(self):
         """
@@ -47,8 +49,8 @@ def rotate(self):
         for key in self.images.keys():
             try:
                 self.images[key] = ndimage.rotate(self.images[key], self.rotation_angle, reshape=False)
-            except:
-                print(key + ' not available')
+            except Exception as e:
+                print(e)
 
     def elastic_transform(self, random_state=None):
         """
@@ -78,5 +80,5 @@ def elastic_transform(self, random_state=None):
         for key in self.images.keys():
             try:
                 self.images[key] = cv2.warpAffine(self.images[key], M, shape_size[::-1], borderMode=cv2.BORDER_REFLECT_101)
-            except:
-                print(key + ' not available')
+            except Exception as e:
+                print(e)

Image_Processing/Image_Processing_Helper_Functions.py

@@ -87,7 +87,7 @@ def create_training_testing_dataset_from_given_directory(input_dir, output_dir,
             cv2.imwrite(os.path.join(all_dirs[i], filename), all_images[filename])
 
 
-def augment_set(input_dir, output_dir, aug_no=9, modality_types=['hematoxylin', 'CD3', 'PanCK'], tile_size=512):
+def augment_set(input_dir, output_dir, aug_no=9, modality_types=['hematoxylin', 'CD3', 'PanCK']):
     """
     This function augments a co-aligned dataset.
 
@@ -105,20 +105,29 @@ def augment_set(input_dir, output_dir, aug_no=9, modality_types=['hematoxylin',
     """
     if not os.path.exists(output_dir):
         os.makedirs(output_dir)
-    images = os.listdir(input_dir)
-    for img in images:
+    images_original = os.listdir(input_dir)
+    print(f'{len(images_original)} images found')
+
+    count = 0
+    for i,img in enumerate(images_original):
         augmented = 0
         while augmented < aug_no:
             images = {}
             image = cv2.imread(os.path.join(input_dir, img))
+            if i == 0:
+                tile_size = image.shape[0]
+            assert image.shape[1] >= len(modality_types) * tile_size, f'image width ({image.shape[1]}) is not enough for {len(modality_types)} modalities with tile size {tile_size}'
             for i in range(0, len(modality_types)):
                 images[modality_types[i]] = image[:, i * tile_size: (i + 1) * tile_size]
             new_images = images.copy()
-            aug = Augmentation(new_images)
+            aug = Augmentation(new_images, tile_size)
             aug.pipeline()
             cv2.imwrite(os.path.join(output_dir, img.replace('.png', '_' + str(augmented) + '.png')),
                         np.concatenate(list(new_images.values()), 1))
             augmented += 1
+        count += 1
+        if count % 10 == 0 or count == len(images_original):
+             print(f'Done {count}/{len(images_original)}')
 
 
 def augment_created_dataset(input_dir, output_dir, aug_no=9, modality_types=['hematoxylin', 'CD3', 'PanCK'], tile_size=512):