app.py

import numpy as np
import gradio as gr
import cv2

from models.HybridGNet2IGSC import Hybrid
from utils.utils import scipy_to_torch_sparse, genMatrixesLungsHeart
import scipy.sparse as sp
import torch

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
hybrid = None


def getDenseMask(landmarks, h, w):
    RL = landmarks[0:44]
    LL = landmarks[44:94]
    H = landmarks[94:]

    img = np.zeros([h, w], dtype='uint8')

    RL = RL.reshape(-1, 1, 2).astype('int')
    LL = LL.reshape(-1, 1, 2).astype('int')
    H = H.reshape(-1, 1, 2).astype('int')

    img = cv2.drawContours(img, [RL], -1, 1, -1)
    img = cv2.drawContours(img, [LL], -1, 1, -1)
    img = cv2.drawContours(img, [H], -1, 2, -1)

    return img


def getMasks(landmarks, h, w):
    RL = landmarks[0:44]
    LL = landmarks[44:94]
    H = landmarks[94:]

    RL = RL.reshape(-1, 1, 2).astype('int')
    LL = LL.reshape(-1, 1, 2).astype('int')
    H = H.reshape(-1, 1, 2).astype('int')

    RL_mask = np.zeros([h, w], dtype='uint8')
    LL_mask = np.zeros([h, w], dtype='uint8')
    H_mask = np.zeros([h, w], dtype='uint8')

    RL_mask = cv2.drawContours(RL_mask, [RL], -1, 255, -1)
    LL_mask = cv2.drawContours(LL_mask, [LL], -1, 255, -1)
    H_mask = cv2.drawContours(H_mask, [H], -1, 255, -1)

    return RL_mask, LL_mask, H_mask


def calculate_image_tilt(landmarks):
    """Calculate image tilt angle based on lung symmetry"""
    RL = landmarks[0:44]  # Right lung
    LL = landmarks[44:94]  # Left lung
    
    # Find the topmost points of both lungs
    rl_top_idx = np.argmin(RL[:, 1])
    ll_top_idx = np.argmin(LL[:, 1])
    
    rl_top = RL[rl_top_idx]
    ll_top = LL[ll_top_idx]
    
    # Calculate angle between the line connecting lung tops and horizontal
    dx = ll_top[0] - rl_top[0]
    dy = ll_top[1] - rl_top[1]
    
    angle_rad = np.arctan2(dy, dx)
    angle_deg = np.degrees(angle_rad)
    
    return angle_deg, rl_top, ll_top

def rotate_points(points, angle_deg, center):
    """Rotate points around a center by given angle"""
    angle_rad = np.radians(-angle_deg)  # Negative to correct the tilt
    cos_a = np.cos(angle_rad)
    sin_a = np.sin(angle_rad)
    
    # Translate to origin
    translated = points - center
    
    # Rotate
    rotated = np.zeros_like(translated)
    rotated[:, 0] = translated[:, 0] * cos_a - translated[:, 1] * sin_a
    rotated[:, 1] = translated[:, 0] * sin_a + translated[:, 1] * cos_a
    
    # Translate back
    return rotated + center

def drawOnTop(img, landmarks, original_shape):
    h, w = original_shape
    output = getDenseMask(landmarks, h, w)

    image = np.zeros([h, w, 3])
    image[:, :, 0] = img + 0.3 * (output == 1).astype('float') - 0.1 * (output == 2).astype('float')
    image[:, :, 1] = img + 0.3 * (output == 2).astype('float') - 0.1 * (output == 1).astype('float')
    image[:, :, 2] = img - 0.1 * (output == 1).astype('float') - 0.2 * (output == 2).astype('float')

    image = np.clip(image, 0, 1)

    RL, LL, H = landmarks[0:44], landmarks[44:94], landmarks[94:]

    # Calculate image tilt and correct it for measurements
    tilt_angle, rl_top, ll_top = calculate_image_tilt(landmarks)
    image_center = np.array([w/2, h/2])
    
    # Draw tilt reference line (green)
    image = cv2.line(image, (int(rl_top[0]), int(rl_top[1])), (int(ll_top[0]), int(ll_top[1])), (0, 1, 0), 1)
    
    # Add tilt angle text
    tilt_text = f"Tilt: {tilt_angle:.1f}°"
    cv2.putText(image, tilt_text, (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (0, 1, 0), 2)
    
    # Correct landmarks for tilt
    if abs(tilt_angle) > 2:  # Only correct if tilt is significant
        RL_corrected = rotate_points(RL, tilt_angle, image_center)
        LL_corrected = rotate_points(LL, tilt_angle, image_center)
        H_corrected = rotate_points(H, tilt_angle, image_center)
        cv2.putText(image, "Tilt Corrected", (10, 60), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (1, 1, 0), 2)
    else:
        RL_corrected, LL_corrected, H_corrected = RL, LL, H

    # Draw the landmarks as dots
    for l in RL:
        image = cv2.circle(image, (int(l[0]), int(l[1])), 5, (1, 0, 1), -1)
    for l in LL:
        image = cv2.circle(image, (int(l[0]), int(l[1])), 5, (1, 0, 1), -1)
    for l in H:
        image = cv2.circle(image, (int(l[0]), int(l[1])), 5, (1, 1, 0), -1)

    # Draw horizontal lines for CTR calculation using corrected landmarks
    # Heart (red line) - using corrected coordinates
    heart_xmin = int(np.min(H_corrected[:, 0]))
    heart_xmax = int(np.max(H_corrected[:, 0]))
    heart_y = int(np.mean([H_corrected[np.argmin(H_corrected[:, 0]), 1], H_corrected[np.argmax(H_corrected[:, 0]), 1]]))
    image = cv2.line(image, (heart_xmin, heart_y), (heart_xmax, heart_y), (1, 0, 0), 2)
    
    # Add vertical lines at heart endpoints to verify widest points
    line_length = 30  # Length of vertical indicator lines
    image = cv2.line(image, (heart_xmin, heart_y - line_length), (heart_xmin, heart_y + line_length), (1, 0, 0), 2)
    image = cv2.line(image, (heart_xmax, heart_y - line_length), (heart_xmax, heart_y + line_length), (1, 0, 0), 2)

    # Thorax (blue line) - using corrected coordinates
    thorax_xmin = int(min(np.min(RL_corrected[:, 0]), np.min(LL_corrected[:, 0])))
    thorax_xmax = int(max(np.max(RL_corrected[:, 0]), np.max(LL_corrected[:, 0])))
    # Find y at leftmost and rightmost points
    if np.min(RL_corrected[:, 0]) < np.min(LL_corrected[:, 0]):
        thorax_ymin = RL_corrected[np.argmin(RL_corrected[:, 0]), 1]
    else:
        thorax_ymin = LL_corrected[np.argmin(LL_corrected[:, 0]), 1]
    if np.max(RL_corrected[:, 0]) > np.max(LL_corrected[:, 0]):
        thorax_ymax = RL_corrected[np.argmax(RL_corrected[:, 0]), 1]
    else:
        thorax_ymax = LL_corrected[np.argmax(LL_corrected[:, 0]), 1]
    thorax_y = int(np.mean([thorax_ymin, thorax_ymax]))
    image = cv2.line(image, (thorax_xmin, thorax_y), (thorax_xmax, thorax_y), (0, 0, 1), 2)
    
    # Add vertical lines at thorax endpoints to verify widest points
    image = cv2.line(image, (thorax_xmin, thorax_y - line_length), (thorax_xmin, thorax_y + line_length), (0, 0, 1), 2)
    image = cv2.line(image, (thorax_xmax, thorax_y - line_length), (thorax_xmax, thorax_y + line_length), (0, 0, 1), 2)
    
    # Store corrected landmarks for CTR calculation
    return image, (RL_corrected, LL_corrected, H_corrected, tilt_angle)

    return image


def loadModel(device):
    A, AD, D, U = genMatrixesLungsHeart()
    N1 = A.shape[0]
    N2 = AD.shape[0]

    A = sp.csc_matrix(A).tocoo()
    AD = sp.csc_matrix(AD).tocoo()
    D = sp.csc_matrix(D).tocoo()
    U = sp.csc_matrix(U).tocoo()

    D_ = [D.copy()]
    U_ = [U.copy()]

    config = {}

    config['n_nodes'] = [N1, N1, N1, N2, N2, N2]
    A_ = [A.copy(), A.copy(), A.copy(), AD.copy(), AD.copy(), AD.copy()]

    A_t, D_t, U_t = ([scipy_to_torch_sparse(x).to(device) for x in X] for X in (A_, D_, U_))

    config['latents'] = 64
    config['inputsize'] = 1024

    f = 32
    config['filters'] = [2, f, f, f, f // 2, f // 2, f // 2]
    config['skip_features'] = f

    hybrid = Hybrid(config.copy(), D_t, U_t, A_t).to(device)
    hybrid.load_state_dict(torch.load("weights/weights.pt", map_location=torch.device(device)))
    hybrid.eval()

    return hybrid


def pad_to_square(img):
    h, w = img.shape[:2]

    if h > w:
        padw = (h - w)
        auxw = padw % 2
        img = np.pad(img, ((0, 0), (padw // 2, padw // 2 + auxw)), 'constant')

        padh = 0
        auxh = 0

    else:
        padh = (w - h)
        auxh = padh % 2
        img = np.pad(img, ((padh // 2, padh // 2 + auxh), (0, 0)), 'constant')

        padw = 0
        auxw = 0

    return img, (padh, padw, auxh, auxw)


def preprocess(input_img):
    img, padding = pad_to_square(input_img)

    h, w = img.shape[:2]
    if h != 1024 or w != 1024:
        img = cv2.resize(img, (1024, 1024), interpolation=cv2.INTER_CUBIC)

    return img, (h, w, padding)


def removePreprocess(output, info):
    h, w, padding = info

    if h != 1024 or w != 1024:
        output = output * h
    else:
        output = output * 1024

    padh, padw, auxh, auxw = padding

    output[:, 0] = output[:, 0] - padw // 2
    output[:, 1] = output[:, 1] - padh // 2

    return output


def calculate_ctr(landmarks, corrected_landmarks=None):
    if corrected_landmarks is not None:
        RL, LL, H, tilt_angle = corrected_landmarks
    else:
        H = landmarks[94:]
        RL = landmarks[0:44]
        LL = landmarks[44:94]
        tilt_angle = 0
    
    cardiac_width = np.max(H[:, 0]) - np.min(H[:, 0])
    thoracic_width = max(np.max(RL[:, 0]), np.max(LL[:, 0])) - min(np.min(RL[:, 0]), np.min(LL[:, 0]))
    ctr = cardiac_width / thoracic_width if thoracic_width > 0 else 0
    return round(ctr, 3), abs(tilt_angle)


def segment(input_img):
    global hybrid, device

    if hybrid is None:
        hybrid = loadModel(device)

    input_img = cv2.imread(input_img, 0) / 255.0
    original_shape = input_img.shape[:2]

    img, (h, w, padding) = preprocess(input_img)

    data = torch.from_numpy(img).unsqueeze(0).unsqueeze(0).to(device).float()

    with torch.no_grad():
        output = hybrid(data)[0].cpu().numpy().reshape(-1, 2)

    output = removePreprocess(output, (h, w, padding))

    output = output.astype('int')

    outseg, corrected_data = drawOnTop(input_img, output, original_shape)

    seg_to_save = (outseg.copy() * 255).astype('uint8')
    cv2.imwrite("tmp/overlap_segmentation.png", cv2.cvtColor(seg_to_save, cv2.COLOR_RGB2BGR))

    ctr_value, tilt_angle = calculate_ctr(output, corrected_data)
    
    # Add tilt warning to interpretation
    tilt_warning = ""
    if tilt_angle > 5:
        tilt_warning = f" (⚠️ Image tilted {tilt_angle:.1f}° - measurement corrected)"
    elif tilt_angle > 2:
        tilt_warning = f" (Image tilted {tilt_angle:.1f}° - corrected)"
    
    if ctr_value < 0.5:
        interpretation = f"Normal{tilt_warning}"
    elif 0.51 <= ctr_value <= 0.55:
        interpretation = f"Mild Cardiomegaly (CTR 51-55%){tilt_warning}"
    elif 0.56 <= ctr_value <= 0.60:
        interpretation = f"Moderate Cardiomegaly (CTR 56-60%){tilt_warning}"
    elif ctr_value > 0.60:
        interpretation = f"Severe Cardiomegaly (CTR > 60%){tilt_warning}"
    else:
        interpretation = f"Cardiomegaly{tilt_warning}"

    return outseg, "tmp/overlap_segmentation.png", ctr_value, interpretation


if __name__ == "__main__":
    with gr.Blocks() as demo:
        gr.Markdown("""
                    # Chest X-ray HybridGNet Segmentation.

                    Demo of the HybridGNet model introduced in "Improving anatomical plausibility in medical image segmentation via hybrid graph neural networks: applications to chest x-ray analysis."

                    Instructions:
                    1. Upload a chest X-ray image (PA or AP) in PNG or JPEG format.
                    2. Click on "Segment Image".

                    Note: Pre-processing is not needed, it will be done automatically and removed after the segmentation.

                    Please check citations below.                    
                    """)

        with gr.Tab("Segment Image"):
            with gr.Row():
                with gr.Column():
                    image_input = gr.Image(type="filepath", height=750)

                    with gr.Row():
                        clear_button = gr.Button("Clear")
                        image_button = gr.Button("Segment Image")

                    gr.Examples(inputs=image_input,
                                examples=['utils/example1.jpg', 'utils/example2.jpg', 'utils/example3.png',
                                          'utils/example4.jpg'])

                with gr.Column():
                    image_output = gr.Image(type="filepath", height=750)

                    with gr.Row():
                        ctr_output = gr.Number(label="CTR (Cardiothoracic Ratio)")
                        ctr_interpretation = gr.Textbox(label="Interpretation", interactive=False)

                    results = gr.File()

        gr.Markdown("""
                    If you use this code, please cite:

                    ```
                    @article{gaggion2022TMI,
                        doi = {10.1109/tmi.2022.3224660},
                        url = {https://doi.org/10.1109%2Ftmi.2022.3224660},
                        year = 2022,
                        publisher = {Institute of Electrical and Electronics Engineers ({IEEE})},
                        author = {Nicolas Gaggion and Lucas Mansilla and Candelaria Mosquera and Diego H. Milone and Enzo Ferrante},
                        title = {Improving anatomical plausibility in medical image segmentation via hybrid graph neural networks: applications to chest x-ray analysis},
                        journal = {{IEEE} Transactions on Medical Imaging}
                    }
                    ```

                    This model was trained following the procedure explained on:

                    ```
                    @INPROCEEDINGS{gaggion2022ISBI,
                        author={Gaggion, Nicolás and Vakalopoulou, Maria and Milone, Diego H. and Ferrante, Enzo},
                        booktitle={2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI)}, 
                        title={Multi-Center Anatomical Segmentation with Heterogeneous Labels Via Landmark-Based Models}, 
                        year={2023},
                        volume={},
                        number={},
                        pages={1-5},
                        doi={10.1109/ISBI53787.2023.10230691}
                    }
                    ```

                    Example images extracted from Wikipedia, released under:
                    1. CC0 Universial Public Domain. Source: https://commons.wikimedia.org/wiki/File:Normal_posteroanterior_(PA)_chest_radiograph_(X-ray).jpg
                    2. Creative Commons Attribution-Share Alike 4.0 International. Source: https://commons.wikimedia.org/wiki/File:Chest_X-ray.jpg
                    3. Creative Commons Attribution 3.0 Unported. Source https://commons.wikimedia.org/wiki/File:Implantable_cardioverter_defibrillator_chest_X-ray.jpg
                    4. Creative Commons Attribution-Share Alike 3.0 Unported. Source: https://commons.wikimedia.org/wiki/File:Medical_X-Ray_imaging_PRD06_nevit.jpg

                    Author: Nicolás Gaggion
                    Website: [ngaggion.github.io](https://ngaggion.github.io/)

                    """)

        clear_button.click(lambda: None, None, image_input, queue=False)
        clear_button.click(lambda: None, None, image_output, queue=False)
        clear_button.click(lambda: None, None, ctr_output, queue=False)
        clear_button.click(lambda: None, None, ctr_interpretation, queue=False)

        image_button.click(segment, inputs=image_input, outputs=[image_output, results, ctr_output, ctr_interpretation], queue=False)

    demo.launch()