Test video output

README.md

@@ -1,8 +1,8 @@
 ---
 title: Swinunetr Dicom Video
-emoji: 🏃
-colorFrom: purple
-colorTo: red
+emoji: 📖🎬
+colorFrom: blue
+colorTo: purple
 sdk: gradio
 sdk_version: 3.0.24
 app_file: app.py
@@ -10,4 +10,9 @@ pinned: false
 license: apache-2.0
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+This repository contains the code for UNETR: Transformers for 3D Medical Image Segmentation. UNETR is the first 3D segmentation network that uses a pure vision transformer as its encoder without relying on CNNs for feature extraction. The code presents a volumetric (3D) multi-organ segmentation application using the BTCV challenge dataset.
+
+Check out the Beyond the Cranial Vault source Swin-UNET models [here](https://huggingface.co/darragh/swinunetr-btcv-small). Also in the link, you can see links to the original BTCV winning solution.
+
+This is a small demo on a subset of the test data for the [BTCV competition](https://zenodo.org/record/1169361#.YtGvn-xKhb8).
+

app.py

@@ -0,0 +1,113 @@
+import sys
+import os
+import glob
+import shutil
+import torch
+import argparse
+import mediapy
+import cv2
+import numpy as np
+import gradio as gr 
+from skimage import color, img_as_ubyte
+from monai import transforms, data
+
+os.system("git clone https://github.com/darraghdog/Project-MONAI-research-contributions pmrc")
+sys.path.append("pmrc/SwinUNETR/BTCV")
+from swinunetr import SwinUnetrModelForInference, SwinUnetrConfig
+
+
+ffmpeg_path = shutil.which('ffmpeg')
+mediapy.set_ffmpeg(ffmpeg_path)
+
+model = SwinUnetrModelForInference.from_pretrained('darragh/swinunetr-btcv-tiny')
+model.eval()
+
+input_files = glob.glob('pmrc/SwinUNETR/BTCV/dataset/imagesSampleTs/*.nii.gz')
+input_files = dict((f.split('/')[-1], f) for f in input_files)
+
+# Load and process dicom with monai transforms
+test_transform = transforms.Compose(
+    [
+        transforms.LoadImaged(keys=["image"]),
+        transforms.AddChanneld(keys=["image"]),
+        transforms.Spacingd(keys="image",
+                            pixdim=(1.5, 1.5, 2.0),
+                            mode="bilinear"),
+        transforms.ScaleIntensityRanged(keys=["image"],
+                                        a_min=-175.0,
+                                        a_max=250.0,
+                                        b_min=0.0,
+                                        b_max=1.0,
+                                        clip=True),
+        # transforms.Resized(keys=["image"], spatial_size = (256,256,-1)), 
+        transforms.ToTensord(keys=["image"]),
+    ])
+
+
+
+# Create Data Loader
+def create_dl(test_files):
+    ds = test_transform(test_files)
+    loader = data.DataLoader(ds,
+                             batch_size=1,
+                             shuffle=False)
+    return loader
+
+# Inference and video generation
+def generate_dicom_video(selected_file):
+    
+    test_file = input_files[selected_file]
+    test_files = [{'image': test_file}]
+    dl = create_dl(test_files)
+    batch = next(iter(dl))
+    
+    tst_inputs = batch["image"]
+    tst_inputs = tst_inputs[:,:,:,:,-32:]
+    
+    with torch.no_grad():
+        outputs = model(tst_inputs,
+                            (96,96,96),
+                            8,
+                            overlap=0.5,
+                            mode="gaussian")
+    tst_outputs = torch.softmax(outputs.logits, 1)
+    tst_outputs = torch.argmax(tst_outputs, axis=1)
+    
+    # Write frames to video
+    for inp, outp in zip(tst_inputs, tst_outputs):
+
+        frames = []
+        for idx in range(inp.shape[-1]):
+            # Segmentation
+            seg = outp[:,:,idx].numpy().astype(np.uint8)
+            # Input dicom frame
+            img = (inp[0,:,:,idx]*255).numpy().astype(np.uint8)
+            img = cv2.cvtColor(img,cv2.COLOR_GRAY2RGB)
+            frame = color.label2rgb(seg,img, bg_label = 0)
+            frame = img_as_ubyte(frame)
+            frame = np.concatenate((img, frame), 1)
+            frames.append(frame)
+        mediapy.write_video("dicom.mp4", frames, fps=4)
+    
+    return 'dicom.mp4'
+
+
+'''
+test_file = glob.glob('pmrc/SwinUNETR/BTCV/dataset/imagesSampleTs/*.nii.gz')[0]
+generate_dicom_video(test_file)
+'''
+
+demo = gr.Blocks()
+
+with demo:
+    selected_dicom_key = gr.inputs.Dropdown(
+            choices=sorted(input_files),
+            type="value",
+            label="Select a dicom file")
+    button_gen_video = gr.Button("Generate Video")
+    output_interpolation = gr.Video(label="Generated Video")
+    button_gen_video.click(fn=generate_dicom_video, inputs=selected_dicom_key, outputs=output_interpolation)
+
+demo.launch(debug=True, enable_queue=True)
+
+

packages.txt

@@ -0,0 +1 @@
+ffmpeg

requirements.txt

@@ -0,0 +1,12 @@
+transformers==4.20.1
+torch==1.10.0
+
+git+https://github.com/Project-MONAI/MONAI#egg.gitmonai@0.8.1+271.g07de215c
+nibabel==3.1.1
+tqdm==4.59.0
+einops==0.4.1
+tensorboardX==2.1
+scipy==1.5.0
+mediapy==1.0.3
+scikit-image==0.17.2
+opencv-python==4.6.0.66