CVAT deid (#999)

* Fix network for tooltracking based on monai rc2 + optimize transform

Signed-off-by: Sachidanand Alle <[email protected]>

* fix nan issue for variance scoring + optimize post transform

Signed-off-by: Sachidanand Alle <[email protected]>

* fix comment

Signed-off-by: Sachidanand Alle <[email protected]>

* fix deid logic

Signed-off-by: Sachidanand Alle <[email protected]>

* fix normalize logic for tool

Signed-off-by: Sachidanand Alle <[email protected]>

* fix args

Signed-off-by: Sachidanand Alle <[email protected]>

Signed-off-by: Sachidanand Alle <[email protected]>

Author: SachidanandAlle

Dockerfile

@@ -14,7 +14,7 @@
 # to use different version of MONAI pass `--build-arg MONAI_IMAGE=...`
 # to exclude ORTHANC pass `--build-arg ORTHANC=false`
 
-ARG MONAI_IMAGE=projectmonai/monai:1.0.0rc1
+ARG MONAI_IMAGE=projectmonai/monai:1.0.0rc3
 ARG ORTHANC=false
 ARG BUILD_OHIF=true
 

monailabel/datastore/cvat.py

@@ -16,6 +16,7 @@
 import tempfile
 import time
 
+import numpy as np
 import requests
 from PIL import Image
 from requests.auth import HTTPBasicAuth
@@ -39,6 +40,7 @@ def __init__(
         image_quality=70,
         labels=None,
         normalize_label=True,
+        segment_size=1,
         **kwargs,
     ):
         default_labels = [
@@ -47,15 +49,17 @@ def __init__(
             {"name": "OutBody", "attributes": [], "color": "#0000ff"},
         ]
         labels = labels if labels else default_labels
-        labels = json.loads(labels) if isinstance(labels, str) else self.json()
+        labels = json.loads(labels) if isinstance(labels, str) else labels
 
         self.api_url = api_url.rstrip("/").strip()
         self.auth = HTTPBasicAuth(username, password) if username else None
         self.project = project
         self.task_prefix = task_prefix
         self.image_quality = image_quality
         self.labels = labels
+        self.label_map = {l["name"]: idx for idx, l in enumerate(labels, start=1)}
         self.normalize_label = normalize_label
+        self.segment_size = segment_size
 
         logger.info(f"CVAT:: API URL: {api_url}")
         logger.info(f"CVAT:: UserName: {username}")
@@ -65,6 +69,7 @@ def __init__(
         logger.info(f"CVAT:: Image Quality: {image_quality}")
         logger.info(f"CVAT:: Labels: {labels}")
         logger.info(f"CVAT:: Normalize Label: {normalize_label}")
+        logger.info(f"CVAT:: Segment Size: {normalize_label}")
 
         super().__init__(datastore_path=datastore_path, **kwargs)
 
@@ -117,7 +122,10 @@ def get_cvat_task_id(self, project_id, create):
             task_name = f"{self.task_prefix}_{version}"
             logger.info(f"Creating new CVAT Task: {task_name}; project: {self.project}")
 
-            body = {"name": task_name, "labels": [], "project_id": project_id, "subset": "Train", "segment_size": 1}
+            body = {"name": task_name, "labels": [], "project_id": project_id, "subset": "Train"}
+            if self.segment_size:
+                body["segment_size"] = self.segment_size
+
             task = requests.post(f"{self.api_url}/api/tasks", auth=self.auth, json=body).json()
             logger.debug(task)
             task_id = task["id"]
@@ -157,6 +165,18 @@ def trigger_automation(self, function):
                 r = requests.post(f"{self.api_url}/api/lambda/requests?org=", json=body, auth=self.auth).json()
                 logger.info(r)
 
+    def _load_labelmap_txt(self, file):
+        labelmap = {}
+        if os.path.exists(file):
+            with open(file) as f:
+                for line in f.readlines():
+                    if line and not line.startswith("#"):
+                        fields = line.split(":")
+                        name = fields[0]
+                        rgb = tuple(int(c) for c in fields[1].split(","))
+                        labelmap[name] = rgb
+        return labelmap
+
     def download_from_cvat(self, max_retry_count=5, retry_wait_time=10):
         if self.task_status() != "completed":
             logger.info("No Tasks exists with completed status to refresh/download the final labels")
@@ -190,10 +210,19 @@ def download_from_cvat(self, max_retry_count=5, retry_wait_time=10):
 
                         dest = os.path.join(final_labels, f)
                         if self.normalize_label:
-                            Image.open(label).convert("L").point(lambda x: 0 if x < 128 else 255, "1").save(dest)
+                            img = np.array(Image.open(label))
+                            mask = np.zeros_like(img)
+
+                            labelmap = self._load_labelmap_txt(os.path.join(tmp_folder, "labelmap.txt"))
+                            for name, color in labelmap.items():
+                                if name in self.label_map:
+                                    idx = self.label_map.get(name)
+                                    mask[np.all(img == color, axis=-1)] = idx
+                            Image.fromarray(mask[:, :, 0]).save(dest)  # single channel
+                            logger.info(f"Copy Final Label: {label} to {dest}; unique: {np.unique(mask)}")
                         else:
                             Image.open(label).save(dest)
-                        logger.info(f"Copy Final Label: {label} to {dest}")
+                            logger.info(f"Copy Final Label: {label} to {dest}")
 
                 # Rename after consuming/downloading the labels
                 patch_url = f"{self.api_url}/api/tasks/{task_id}"
@@ -206,3 +235,51 @@ def download_from_cvat(self, max_retry_count=5, retry_wait_time=10):
                 logger.error(f"{retry} => Failed to download...")
             retry_count = retry_count + 1
         return None
+
+
+def main():
+    from pathlib import Path
+
+    from monailabel.config import settings
+
+    settings.MONAI_LABEL_DATASTORE_AUTO_RELOAD = False
+    settings.MONAI_LABEL_DATASTORE_FILE_EXT = ["*.png", "*.jpg", "*.jpeg", ".xml"]
+    settings.MONAI_LABEL_DATASTORE = "cvat"
+    settings.MONAI_LABEL_DATASTORE_URL = "http://10.117.19.88:8080"
+    settings.MONAI_LABEL_DATASTORE_USERNAME = "sachi"
+    settings.MONAI_LABEL_DATASTORE_PASSWORD = "sachi"
+
+    os.putenv("MASTER_ADDR", "127.0.0.1")
+    os.putenv("MASTER_PORT", "1234")
+
+    logging.basicConfig(
+        level=logging.INFO,
+        format="[%(asctime)s] [%(process)s] [%(threadName)s] [%(levelname)s] (%(name)s:%(lineno)d) - %(message)s",
+        datefmt="%Y-%m-%d %H:%M:%S",
+        force=True,
+    )
+
+    home = str(Path.home())
+    studies = f"{home}/Dataset/picked/all"
+
+    ds = CVATDatastore(
+        datastore_path=studies,
+        api_url=settings.MONAI_LABEL_DATASTORE_URL,
+        username=settings.MONAI_LABEL_DATASTORE_USERNAME,
+        password=settings.MONAI_LABEL_DATASTORE_PASSWORD,
+        project="MONAILabel",
+        task_prefix="ActiveLearning_Iteration",
+        image_quality=70,
+        labels=None,
+        normalize_label=True,
+        segment_size=0,
+        extensions=settings.MONAI_LABEL_DATASTORE_FILE_EXT,
+        auto_reload=settings.MONAI_LABEL_DATASTORE_AUTO_RELOAD,
+    )
+    ds.download_from_cvat()
+
+    # studies = f"{home}/Dataset/Holoscan/flattened/images"
+
+
+if __name__ == "__main__":
+    main()

plugins/cvat/endoscopy/deid.yaml

@@ -18,11 +18,12 @@ metadata:
     framework: pytorch
     spec: |
       [
-        { "id": 1, "name": "Tool" }
+        { "id": 1, "name": "InBody" },
+        { "id": 2, "name": "OutBody" }
       ]
 
 spec:
-  description: A pre-trained classification model for Endoscopy to flag if image contains Tool
+  description: A pre-trained classification model for Endoscopy to flag if image follows InBody or OutBody
   runtime: 'python:3.8'
   handler: main:handler
   eventTimeout: 30s

plugins/cvat/main.py

@@ -89,17 +89,25 @@ def handler(context, event):
     prediction = json_data["params"].get("prediction")
     if prediction:
         context.logger.info(f"(Classification) Prediction: {prediction}")
+
+        # CVAT Limitation:: tag is not yet supported https://github.com/opencv/cvat/issues/4212
+        # CVAT Limitation:: select highest score and create bbox to represent as tag
+        e = None
         for element in prediction:
             if element["score"] > 0:
-                # CVAT Limitation:: tag is not yet supported https://github.com/opencv/cvat/issues/4212
-                results.append(
-                    {
-                        "label": element["label"],
-                        "confidence": element["score"],
-                        "type": "rectangle",
-                        "points": [0, 0, image_np.shape[0] - 1, image_np.shape[1] - 1],
-                    }
-                )
+                e = element if e is None or element["score"] > e["score"] else e
+                context.logger.info(f"New Max Element: {e}")
+
+        context.logger.info(f"Final Element with Max Score: {e}")
+        if e:
+            results.append(
+                {
+                    "label": e["label"],
+                    "confidence": e["score"],
+                    "type": "rectangle",
+                    "points": [0, 0, image_np.shape[0] - 1, image_np.shape[1] - 1],
+                }
+            )
         context.logger.info(f"(Classification) Results: {results}")
     else:
         interactor = strtobool(os.environ.get("INTERACTOR_MODEL", "false"))

requirements.txt

@@ -10,7 +10,7 @@
 # limitations under the License.
 
 torch>=1.7
-monai[nibabel, skimage, pillow, tensorboard, gdown, ignite, torchvision, itk, tqdm, lmdb, psutil, openslide, fire]>=1.0.0rc2
+monai[nibabel, skimage, pillow, tensorboard, gdown, ignite, torchvision, itk, tqdm, lmdb, psutil, openslide, fire]>=1.0.0rc3
 uvicorn==0.17.6
 pydantic==1.9.1
 python-dotenv==0.20.0

sample-apps/endoscopy/lib/configs/deid.py

@@ -34,7 +34,7 @@ def init(self, name: str, model_dir: str, conf: Dict[str, str], planner: Any, **
         super().init(name, model_dir, conf, planner, **kwargs)
 
         # Labels
-        self.labels = {"Tool": 0}
+        self.labels = {"InBody": 0, "OutBody": 1}
 
         # Model Files
         self.path = [

sample-apps/endoscopy/lib/infers/deepedit.py

@@ -76,9 +76,7 @@ def post_transforms(self, data=None) -> Sequence[Callable]:
         return [
             EnsureTyped(keys="pred", device=data.get("device") if data else None),
             Activationsd(keys="pred", sigmoid=True),
-            AsDiscreted(keys="pred", threshold_values=True, logit_thresh=0.5),
-            # NormalizeLabeld(keys="pred"),
-            # DumpImagePrediction2Dd(image_path="img.png", pred_path="pred.png"),
+            AsDiscreted(keys="pred", threshold=0.5),
             Restored(keys="pred", ref_image="image"),
             SqueezeDimd(keys="pred"),
             ToNumpyd(keys="pred", dtype=np.uint8),

sample-apps/endoscopy/lib/infers/deid.py

@@ -15,7 +15,7 @@
 import numpy as np
 import torch
 from monai.inferers import Inferer, SimpleInferer
-from monai.transforms import AsChannelFirstd, AsDiscreted, CastToTyped, EnsureTyped, NormalizeIntensityd, Resized
+from monai.transforms import AsChannelFirstd, AsDiscreted, CastToTyped, NormalizeIntensityd, Resized
 
 from monailabel.interfaces.tasks.infer import InferTask, InferType
 from monailabel.transform.pre import LoadImageExd
@@ -57,7 +57,6 @@ def info(self) -> Dict[str, Any]:
     def pre_transforms(self, data=None) -> Sequence[Callable]:
         return [
             LoadImageExd(keys="image", dtype=np.uint8),
-            EnsureTyped(keys="image", device=data.get("device") if data else None),
             AsChannelFirstd(keys="image"),
             Resized(keys="image", spatial_size=(256, 256), mode="bilinear"),
             CastToTyped(keys="image", dtype=torch.float32),
@@ -69,7 +68,6 @@ def inferer(self, data=None) -> Inferer:
 
     def post_transforms(self, data=None) -> Sequence[Callable]:
         return [
-            EnsureTyped(keys="pred", device=data.get("device") if data else None),
             AsDiscreted(keys="pred", argmax=True, to_onehot=2),
         ]
 

sample-apps/endoscopy/lib/trainers/deepedit.py

@@ -106,7 +106,7 @@ def train_pre_transforms(self, context: Context):
     def train_post_transforms(self, context: Context):
         return [
             Activationsd(keys="pred", sigmoid=True),
-            AsDiscreted(keys="pred", threshold_values=True, logit_thresh=0.5),
+            AsDiscreted(keys="pred", threshold=0.5),
         ]
 
     def val_inferer(self, context: Context):

sample-apps/endoscopy/lib/trainers/deid.py

@@ -63,7 +63,7 @@ def loss_function(self, context: Context):
     def train_pre_transforms(self, context: Context):
         return [
             LoadImaged(keys=("image", "label"), dtype=np.uint8),
-            LabelToBinaryClassd(keys="label"),
+            LabelToBinaryClassd(keys="label", offset=2),
             ToTensord(keys=("image", "label")),
             AsChannelFirstd("image"),
             Resized(keys="image", spatial_size=(256, 256), mode="bilinear"),
@@ -84,7 +84,7 @@ def train_post_transforms(self, context: Context):
     def val_pre_transforms(self, context: Context):
         return [
             LoadImaged(keys=("image", "label"), dtype=np.uint8),
-            LabelToBinaryClassd(keys="label"),
+            LabelToBinaryClassd(keys="label", offset=2),
             ToTensord(keys=("image", "label")),
             AsChannelFirstd("image"),
             Resized(keys="image", spatial_size=(256, 256), mode="bilinear"),