app.py

import cv2
import glob
import gradio as gr
import mediapy
import nibabel
import numpy as np
import shutil
import torch
import torch.nn.functional as F

from omegaconf import OmegaConf
from skp import builder


def window(x, WL=400, WW=2500):
    lower, upper = WL - WW // 2, WL + WW // 2
    x = np.clip(x, lower, upper)
    x = x - lower
    x = x / (upper - lower)
    return (x * 255).astype("uint8")


def rescale(x):
    x = x / 255.
    x = x - 0.5
    x = x * 2.0
    return x


def get_cervical_spine_coordinates(x, original_shape):
    # Assumes x is torch tensor with shape (1, 8, Z, H, W)
    x = x.squeeze(0).numpy()[:7]
    rescale_factor = [original_shape[0] / x.shape[1], original_shape[1] / x.shape[2], original_shape[2] / x.shape[3]]
    coords_dict = {}
    for level in range(x.shape[0]):
        coords = np.where(x[level] >= 0.4)
        coords = np.vstack(coords).astype("float")
        coords[0] = coords[0] * rescale_factor[0]
        coords[1] = coords[1] * rescale_factor[1]
        coords[2] = coords[2] * rescale_factor[2]
        coords = coords.astype("int")
        coords_dict[level] = coords[0].min(), coords[0].max(),\
                             coords[1].min(), coords[1].max(),\
                             coords[2].min(), coords[2].max()
    return coords_dict


def generate_segmentation_video(study):
    img = nibabel.load(study).get_fdata()[:, ::-1, ::-1].transpose(2, 1, 0)
    img = window(img)

    X = torch.from_numpy(img).float().unsqueeze(0).unsqueeze(0)
    X = F.interpolate(X, size=(192, 192, 192), mode="nearest")
    X = rescale(X)
    with torch.no_grad():
        seg_output = seg_model(X)

    seg_output = torch.sigmoid(seg_output)
    c_spine_coords = get_cervical_spine_coordinates(seg_output, img.shape)

    chunk_features = []
    for level, coords in c_spine_coords.items():
        z1, z2, h1, h2, w1, w2 = coords
        X = torch.from_numpy(img[z1:z2+1, h1:h2+1, w1:w2+1]).float().unsqueeze(0).unsqueeze(0)
        X = F.interpolate(X, size=(64, 288, 288), mode="nearest")
        X = rescale(X)
        with torch.no_grad():
            chunk_features.append(x3d_model.extract_features(X))

    chunk_features = torch.stack(chunk_features, dim=1)
    with torch.no_grad():
        final_output = torch.sigmoid(seq_model((chunk_features, torch.ones((chunk_features.size(1), )))))

    final_output_dict = {f"C{i+1}": final_output[:, i].item() for i in range(7)}
    final_output_dict["Overall"] = final_output[:, -1].item()

    seg_output = F.interpolate(seg_output, size=img.shape, mode="nearest").squeeze(0).numpy()
    # shape = (8, Z, H, W)
    p_spine = seg_output[:7].sum(0)
    # shape = (Z, H, W)
    seg_output = np.argmax(seg_output[:7], axis=0) + 1
    # shape = (Z, H, W)
    seg_output[p_spine < 0.5] = 0
    seg_output = (seg_output * 255 / 7).astype("uint8")
    seg_output = np.stack([cv2.applyColorMap(_, cv2.COLORMAP_JET) for _ in seg_output])
    seg_output[p_spine < 0.5] = 0

    frames = []
    skip = 8
    for idx in range(0, img.shape[2], skip):
        i = img[:, :, idx]
        o = seg_output[:, :, idx]
        i = cv2.cvtColor(i, cv2.COLOR_GRAY2RGB)
        frame = np.concatenate((i, o), 1)
        frames.append(frame)
    mediapy.write_video("video.mp4", frames, fps=10)
    return "video.mp4", final_output_dict


ffmpeg_path = shutil.which('ffmpeg')
mediapy.set_ffmpeg(ffmpeg_path)

config = OmegaConf.load("configs/pseudoseg000.yaml")
config.model.load_pretrained = "seg.ckpt"
config.model.params.encoder_params.pretrained = False
seg_model = builder.build_model(config).eval()

config = OmegaConf.load("configs/chunk000.yaml")
config.model.load_pretrained = "x3d.ckpt"
config.model.params.pretrained = False
x3d_model = builder.build_model(config).eval()

config = OmegaConf.load("configs/chunkseq003.yaml")
config.model.load_pretrained = "seq.ckpt"
seq_model = builder.build_model(config).eval()

examples = glob.glob("examples/*.nii.gz")

with gr.Blocks(theme="dark-peach") as demo:
    select_study = gr.Dropdown(choices=sorted(examples), type="value", label="Select a study")
    button_predict = gr.Button("Predict")
    video_output = gr.Video(label="Cervical Spine Segmentation")
    label_output = gr.Label(label="Fracture Predictions", show_label=False)
    button_predict.click(fn=generate_segmentation_video,
                         inputs=select_study,
                         outputs=[video_output, label_output])


if __name__ == "__main__":
    demo.launch(debug=True)