Merge pull request #368 from Project-MONAI/pin_integration_tweaks

CPBridge · web-flow · commit f7176c0b6809 · 2022-10-19T15:29:22.000Z
Various fixes to PIN integration example
diff --git a/integrations/nuance_pin/README.md b/integrations/nuance_pin/README.md
@@ -8,7 +8,7 @@ with minimal code changes.
 
 ## Prerequisites
 
-Before setting up and running the example MONAI spleen segmentation app to run as a Nuance PIN App, the user will need to install/download the following libraries.
+Before setting up and running the example MONAI lung nodule detection app to run as a Nuance PIN App, the user will need to install/download the following libraries.
 It is optional to use a GPU for the example app, however, it is recommended that a GPU is used for inference as it is very computationally intensive.
 
 Minimum software requirements:
@@ -30,9 +30,8 @@ cd integrations/nuance_pin
 In this folder you will see the following directory structure
 ```bash
 nuance_pin
-    ├── app                     # directory with MONAI app code
-    ├── lib                     # directory where we will place Nuance PIN wheels
-    ├── model                   # directory where we will place the model used by our MONAI app
+    ├── app/                    # directory with MONAI app code ├── lib/                    # you should create this directory where we will place Nuance PIN wheels
+    ├── model/                  # directory where we will place the model used by our MONAI app
     ├── app_wrapper.py          # Nuance PIN wrapper code
     ├── docker-compose.yml      # docker compose runtime script
     ├── Dockerfile              # container image build script
@@ -48,7 +47,7 @@ To download the test data you may follow the instructions in the [Lund Nodule De
 
 ### Download Nuance PIN SDK
 
-Place the Nuance PIN `ai_service` wheel in the `nuance_pin/lib` folder. This can be obtained in the link provided in step 3 of of the [prerequisites](#prerequisites).
+Place the Nuance PIN `ai_service` wheel in the `nuance_pin/lib` folder. This can be obtained in the link provided in step 4 of of the [prerequisites](#prerequisites).
 
 ### Running the Example App in the Container
 
@@ -57,7 +56,7 @@ Now we are ready to build and start the container that runs our MONAI app as a N
 docker-compose up --build
 ```
 
-If the build is successful the a service will start on `localhost:5000`. We can verify the service is running
+If the build is successful the service will start on `localhost:5000`. We can verify the service is running
 by issuing a "live" request such as
 ```bash
 curl -v http://localhost:5000/aiservice/2/live && echo ""
@@ -136,12 +135,12 @@ python -m AiSvcTest -i ~/Downloads/dcm -o ~/Downloads/dcm/out -s http://localhos
 
 ### Bring Your Own MONAI App
 
-This example integration may be modified to fit any existing MONAI app, however, there may be caveats.
+This example integration may be modified to fit any existing MONAI app by tailoring the files within the `app/` directory, however, there may be caveats.
 
 Nuance PIN requires all artifacts present in the output folder to be also added into the `resultManifest.json` output file
 to consider the run successful. To see what this means in practical terms, check the `resultManifest.json` output from the
 example app we ran the in previous sections. You will notice an entry in `resultManifest.json` that corresponds to the DICOM
-SEG output generated by the underlying MONAI app
+GSPS output generated by the underlying MONAI app
 ```json
   "study": {
     "uid": "1.2.826.0.1.3680043.2.1125.1.67295333199898911264201812221946213",
@@ -153,17 +152,16 @@ SEG output generated by the underlying MONAI app
           {
             "documentType": "application/dicom",
             "groupCode": "default",
-            "name": "dicom_seg-DICOMSEG.dcm",
+            "name": "gsps.dcm",
             "trackingUids": []
           }
         ]
       }
     ]
   },
 ```
-This entry is generated by `app_wrapper.py`, which takes care of adding any DICOM present in the output folder in the `resultManifest.json`
-to ensure that existing MONAI apps complete successfully when deployed in Nuance. In general, however, the developer may need to tailor some
-of the code in `app_wrapper.py` to provide more insight to Nuance's network, such as adding findings, conclusions, etc. and generating more insight
+This entry is generated automatically by Nuance's `ai_service` library as a result of uploading the DICOM GSPS object in `app/post_inference_ops.py`.
+In general, however, the developer may need to tailor some of the code in `app_wrapper.py` to provide more insight to Nuance's network, such as adding findings, conclusions, etc. and generating more insight
 using SNOMED codes. All of this is handled within the Nuance PIN SDK libraries - for more information please consult Nuance PIN [documentation](https://www.nuance.com/healthcare/diagnostics-solutions/precision-imaging-network.html).
 
 In simpler cases, the developer will need to place their code and model under `nuance_pin`. Placing the model under `model` is optional as the model may be placed
diff --git a/integrations/nuance_pin/app/inference.py b/integrations/nuance_pin/app/inference.py
@@ -177,7 +177,7 @@ def pre_process(self, img_reader) -> Compose:
                     keys=[image_key, f"{image_key}_meta_dict"],
                     names=[orig_image_key, f"{orig_image_key}_meta_dict"],
                 ),
-                ToDeviced(keys=image_key, device="cuda"),
+                ToDeviced(keys=image_key, device=self.device),
                 EnsureChannelFirstd(keys=image_key),
                 Spacingd(keys=image_key, pixdim=(0.703125, 0.703125, 1.25)),
                 Orientationd(
diff --git a/integrations/nuance_pin/app/post_inference_ops.py b/integrations/nuance_pin/app/post_inference_ops.py
@@ -46,8 +46,23 @@ def compute(self, op_input: InputContext, op_output: OutputContext, context: Exe
         series_uid = hd.UID()
         series_number = randint(1, 100000)
 
+        # One graphic layer to contain all detections
+        layer = hd.pr.GraphicLayer(
+            layer_name="LUNG_NODULES",
+            order=1,
+            description="Lung Nodule Detections",
+        )
+
+        annotations = []
+
+        all_ref_images = [ins.get_native_sop_instance() for ins in selected_series.series.get_sop_instances()]
+        accession = all_ref_images[0].AccessionNumber
+
         for inst_num, (box_data, box_score) in enumerate(zip(detection_result.box_data, detection_result.score_data)):
 
+            tracking_id = f"{accession}_nodule_{inst_num}"  # site-specific ID
+            tracking_uid = hd.UID()
+
             polyline = hd.pr.GraphicObject(
                 graphic_type=hd.pr.GraphicTypeValues.POLYLINE,
                 graphic_data=np.array(
@@ -60,8 +75,8 @@ def compute(self, op_input: InputContext, op_output: OutputContext, context: Exe
                     ]
                 ),  # coordinates of polyline vertices
                 units=hd.pr.AnnotationUnitsValues.PIXEL,  # units for graphic data
-                tracking_id="lung_nodule_MONAI",  # site-specific ID
-                tracking_uid=hd.UID(),  # highdicom will generate a unique ID
+                tracking_id=tracking_id,
+                tracking_uid=tracking_uid,
             )
 
             self.logger.info(f"Box: {[box_data[0], box_data[1], box_data[3], box_data[4]]}")
@@ -70,14 +85,8 @@ def compute(self, op_input: InputContext, op_output: OutputContext, context: Exe
                 text_value=f"{box_score:.2f}",
                 bounding_box=(box_data[0], box_data[1], box_data[3], box_data[4]),  # left, top, right, bottom
                 units=hd.pr.AnnotationUnitsValues.PIXEL,  # units for bounding box
-                tracking_id="LungNoduleMONAI",  # site-specific ID
-                tracking_uid=hd.UID(),  # highdicom will generate a unique ID
-            )
-
-            layer = hd.pr.GraphicLayer(
-                layer_name="LUNG_NODULE",
-                order=1,
-                description="Lung Nodule Detection",
+                tracking_id=tracking_id,
+                tracking_uid=tracking_uid,
             )
 
             affected_slice_idx = [
@@ -103,37 +112,40 @@ def compute(self, op_input: InputContext, op_output: OutputContext, context: Exe
                 graphic_objects=[polyline],
             )
 
-            # Assemble the components into a GSPS object
-            gsps = hd.pr.GrayscaleSoftcopyPresentationState(
-                referenced_images=ref_images,
-                series_instance_uid=series_uid,
-                series_number=series_number,
-                sop_instance_uid=hd.UID(),
-                instance_number=inst_num + 1,
-                manufacturer="MONAI",
-                manufacturer_model_name="lung_nodule_ct_detection",
-                software_versions="v0.2.0",
-                device_serial_number="",
-                content_label="ANNOTATIONS",
-                graphic_layers=[layer],
-                graphic_annotations=[annotation],
-                institution_name="MONAI",
-                institutional_department_name="Deploy",
-                voi_lut_transformations=[
-                    hd.pr.SoftcopyVOILUTTransformation(
-                        window_center=-550.0,
-                        window_width=1350.0,
-                    )
-                ],
-            )
+            annotations.append(annotation)
+
+        # Assemble the components into a GSPS object
+        gsps = hd.pr.GrayscaleSoftcopyPresentationState(
+            referenced_images=all_ref_images,
+            series_instance_uid=series_uid,
+            series_number=series_number,
+            sop_instance_uid=hd.UID(),
+            instance_number=1,
+            manufacturer="MONAI",
+            manufacturer_model_name="lung_nodule_ct_detection",
+            software_versions="v0.2.0",
+            device_serial_number="",
+            content_label="ANNOTATIONS",
+            graphic_layers=[layer],
+            graphic_annotations=annotations,
+            institution_name="MONAI",
+            institutional_department_name="Deploy",
+            voi_lut_transformations=[
+                hd.pr.SoftcopyVOILUTTransformation(
+                    window_center=-550.0,
+                    window_width=1350.0,
+                )
+            ],
+        )
 
-            gsps.save_as(os.path.join(output_path, f"gsps-{inst_num:04d}.dcm"))
+        gsps_filename = os.path.join(output_path, "gsps.dcm")
+        gsps.save_as(gsps_filename)
 
-            self.upload_gsps(
-                file=os.path.join(output_path, f"gsps-{inst_num:04d}.dcm"),
-                document_detail="MONAI Lung Nodule Detection v0.2.0",
-                series_uid=series_uid,
-            )
+        self.upload_gsps(
+            file=gsps_filename,
+            document_detail="MONAI Lung Nodule Detection v0.2.0",
+            series_uid=series_uid,
+        )
 
 
 @md.input("original_dicom", List[StudySelectedSeries], IOType.IN_MEMORY)
diff --git a/integrations/nuance_pin/app_wrapper.py b/integrations/nuance_pin/app_wrapper.py
@@ -72,9 +72,7 @@ def initialize_class(cls):
         monai_app_class_module = cls.monai_app_module.rsplit(".", 1)[0]
         monai_app_class_name = cls.monai_app_module.rsplit(".", 1)[1]
         if not cls.monai_app_module:
-            raise ValueError(
-                "MONAI App to be run has not been specificed in `MONAI_APP_CLASSPATH` environment variable"
-            )
+            raise ValueError("MONAI App to be run has not been specified in `MONAI_APP_CLASSPATH` environment variable")
 
         monai_app_class = getattr(import_module(monai_app_class_module), monai_app_class_name)
         if monai_app_class is None:
