app.py

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)

# Load model
model = SwinUnetrModelForInference.from_pretrained('darragh/swinunetr-btcv-tiny')
model.eval()

# Pull files from github
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, n_frames):
    
    # Data processor
    test_file = input_files[selected_file]
    test_files = [{'image': test_file}]
    dl = create_dl(test_files)
    batch = next(iter(dl))
    
    # Select dicom slices
    tst_inputs = batch["image"]
    tst_inputs = tst_inputs[:,:,:,:,-n_frames:]

    # Inference    
    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'


theme = 'dark-peach'
with gr.Blocks(theme=theme) as demo:

    gr.Markdown('''<center><h1>SwinUnetr BTCV</h1></center>
	This is a Gradio Blocks app of the winning transformer in the Beyond the Cranial Vault (BTCV) Segmentation Challenge, <a href="https://github.com/darraghdog/Project-MONAI-research-contributions/tree/main/SwinUNETR/BTCV">SwinUnetr</a> (tiny version).
	''')
    selected_dicom_key = gr.inputs.Dropdown(
            choices=sorted(input_files),
            type="value",
            label="Select a dicom file")
    n_frames = gr.Slider(1, 100, value=32, label="Choose the number of dicom slices to process", step = 1)
    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, n_frames], 
                           outputs=output_interpolation)

demo.launch(debug=True, enable_queue=True)