Added a basic transform as discussed in #21

- More transforms can be experimented with later, and the format will essentially remain the same.

capstone/notebooks/miccai_batch_exploration.ipynb

@@ -360,17 +360,22 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 10,
+   "execution_count": 3,
    "metadata": {},
    "outputs": [],
    "source": [
     "def window_stats(patient, window, crop=True):\n",
     "    temp_patient = patient\n",
-    "    window_width, window_level = window\n",
     "    if crop:\n",
     "        temp_patient.crop_data()\n",
-    "        \n",
     "    vol = temp_patient.image.as_numpy()\n",
+    "    \n",
+    "    if window is None:\n",
+    "        window_width = np.math.ceil(np.max(vol) - np.min(vol))\n",
+    "        window_level = np.math.ceil(np.max(vol) - (window_width / 2))\n",
+    "    else:\n",
+    "        window_width, window_level = window\n",
+    "    \n",
     "    clipped = apply_window(vol, window_width, window_level, shift=True)\n",
     "    N = np.prod(vol.shape)\n",
     "    \n",
@@ -379,16 +384,17 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 11,
+   "execution_count": 4,
    "metadata": {},
    "outputs": [
     {
      "name": "stderr",
      "output_type": "stream",
      "text": [
-      "100%|██████████████████████████████████████████████████████████████████████████████████| 25/25 [01:09<00:00,  2.78s/it]\n",
-      "100%|██████████████████████████████████████████████████████████████████████████████████| 25/25 [01:12<00:00,  2.88s/it]\n",
-      "100%|██████████████████████████████████████████████████████████████████████████████████| 25/25 [01:08<00:00,  2.76s/it]\n"
+      "100%|██████████████████████████████████████████████████████████████████████████████████| 25/25 [01:11<00:00,  2.86s/it]\n",
+      "100%|██████████████████████████████████████████████████████████████████████████████████| 25/25 [01:06<00:00,  2.67s/it]\n",
+      "100%|██████████████████████████████████████████████████████████████████████████████████| 25/25 [01:08<00:00,  2.73s/it]\n",
+      "100%|██████████████████████████████████████████████████████████████████████████████████| 25/25 [01:14<00:00,  2.96s/it]\n"
      ]
     }
    ],
@@ -400,12 +406,15 @@
     "results_tissue = patient_collection.apply_function(window_stats, window=WINDOWING_CONFIG[\"soft_tissue\"])\n",
     "\n",
     "# Bone windowing\n",
-    "results_bone = patient_collection.apply_function(window_stats, window=WINDOWING_CONFIG[\"bone\"])"
+    "results_bone = patient_collection.apply_function(window_stats, window=WINDOWING_CONFIG[\"bone\"])\n",
+    "\n",
+    "# No windowing\n",
+    "results_none = patient_collection.apply_function(window_stats, window=None)"
    ]
   },
   {
    "cell_type": "code",
-   "execution_count": 12,
+   "execution_count": 5,
    "metadata": {},
    "outputs": [
     {
@@ -414,7 +423,8 @@
      "text": [
       "Brain windowing stats: mean=0.10677633102763491, std=0.2708013287518585\n",
       "Soft tissue windowing stats: mean=0.13461463937176568, std=0.2668329203556111\n",
-      "Bone windowing stats: mean=0.08499418334609993, std=0.1520881280521281\n"
+      "Bone windowing stats: mean=0.08499418334609993, std=0.1520881280521281\n",
+      "No windowing stats: mean=0.07861779334888626, std=0.13267967531876107\n"
      ]
     }
    ],
@@ -438,7 +448,14 @@
     "sum_vars = temp[:, 1].sum()\n",
     "sum_voxels = temp[:, 2].sum()\n",
     "\n",
-    "print(f\"Bone windowing stats: mean={sum_means / sum_voxels}, std={np.sqrt(sum_vars / sum_voxels)}\")"
+    "print(f\"Bone windowing stats: mean={sum_means / sum_voxels}, std={np.sqrt(sum_vars / sum_voxels)}\")\n",
+    "\n",
+    "temp = np.array(list(results_none.values()))\n",
+    "sum_means = temp[:, 0].sum()\n",
+    "sum_vars = temp[:, 1].sum()\n",
+    "sum_voxels = temp[:, 2].sum()\n",
+    "\n",
+    "print(f\"No windowing stats: mean={sum_means / sum_voxels}, std={np.sqrt(sum_vars / sum_voxels)}\")"
    ]
   },
   {

capstone/transforms/predefined.py

@@ -0,0 +1,40 @@
+import albumentations as A
+from albumentations.pytorch.transforms import ToTensorV2
+from capstone.transforms.transforms_2d import WindowedChannels
+
+_stacked_window_stats = {"mean": (0.107, 0.135, 0.085), "std": (0.271, 0.267, 0.152)}
+# _no_window_stats = {"mean": (0.077), "std": (0.133)}
+
+_minimal_windowed_test_transform = A.Compose(
+    [
+        WindowedChannels(),
+        A.Resize(256, 256),
+        A.Normalize(
+            mean=_stacked_window_stats["mean"],
+            std=_stacked_window_stats["std"],
+            max_pixel_value=1.0,
+        ),
+        ToTensorV2(),
+    ]
+)
+
+
+def minimal_windowed_transforms(split="train"):
+    assert split in ["train", "valid", "test"], "Invalid data split passed"
+    if split == "train":
+        return A.Compose(
+            [
+                WindowedChannels(),
+                A.RandomCrop(256, 256),
+                A.RandomRotate90(),
+                A.HorizontalFlip(),
+                A.Normalize(
+                    mean=_stacked_window_stats["mean"],
+                    std=_stacked_window_stats["std"],
+                    max_pixel_value=1.0,
+                ),
+                ToTensorV2(),
+            ]
+        )
+    else:
+        return _minimal_windowed_test_transform