app.py

import streamlit as st
import cv2
import numpy as np
import pydicom
import tensorflow as tf
import keras
from pydicom.dataset import Dataset, FileDataset
from pydicom.uid import generate_uid
from google.cloud import storage
import os
import io
from PIL import Image
import uuid
import pandas as pd
import tensorflow as tf
from datetime import datetime
from tensorflow import image
from tensorflow.python.keras.models import load_model
from pydicom.pixel_data_handlers.util import apply_voi_lut

# Environment Configuration
os.environ['GOOGLE_APPLICATION_CREDENTIALS'] = "./da-kalbe-63ee33c9cdbb.json"
bucket_name = "da-kalbe-ml-result-png"
storage_client = storage.Client()
bucket_result = storage_client.bucket(bucket_name)
bucket_name_load = "da-ml-models"
bucket_load = storage_client.bucket(bucket_name_load)

model_path = os.path.join("model.h5")
model = tf.keras.models.load_model(model_path)

H, W = 512, 512

test_samples_folder = 'object_detection_test_samples'

def cal_iou(y_true, y_pred):
    x1 = max(y_true[0], y_pred[0])
    y1 = max(y_true[1], y_pred[1])
    x2 = min(y_true[2], y_pred[2])
    y2 = min(y_true[3], y_pred[3])

    intersection_area = max(0, x2 - x1 + 1) * max(0, y2 - y1 + 1)

    true_area = (y_true[2] - y_true[0] + 1) * (y_true[3] - y_true[1] + 1)
    bbox_area = (y_pred[2] - y_pred[0] + 1) * (y_pred[3] - y_pred[1] + 1)

    iou = intersection_area / float(true_area + bbox_area - intersection_area)
    return iou

df = pd.read_excel('BBox_List_2017.xlsx')
labels_dict = dict(zip(df['Image Index'], df['Finding Label']))

def predict(image):
    H, W = 512, 512

    image_resized = cv2.resize(image, (W, H))
    image_normalized = (image_resized - 127.5) / 127.5
    image_normalized = np.expand_dims(image_normalized, axis=0)

    # Prediction
    pred_bbox = model.predict(image_normalized, verbose=0)[0]

    # Rescale the bbox points
    pred_x1 = int(pred_bbox[0] * image.shape[1])
    pred_y1 = int(pred_bbox[1] * image.shape[0])
    pred_x2 = int(pred_bbox[2] * image.shape[1])
    pred_y2 = int(pred_bbox[3] * image.shape[0])

    return (pred_x1, pred_y1, pred_x2, pred_y2)

st.title("AI Integration for Chest X-Ray Imaging")

# Concept 1: Select from test samples
# st.header("Select Test Sample Images")
# test_sample_images = [os.path.join(test_samples_folder, f) for f in os.listdir(test_samples_folder) if f.endswith('.jpg') or f.endswith('.png')]
# test_sample_selected = st.selectbox("Select a test sample image", test_sample_images)
# if test_sample_selected:
#     st.image(test_sample_selected, caption='Selected Test Sample Image', use_column_width=True)
    

# Utility Functions
def upload_to_gcs(image_data: io.BytesIO, filename: str, content_type='application/dicom'):
    """Uploads an image to Google Cloud Storage."""
    try:
        blob = bucket_result.blob(filename)
        blob.upload_from_file(image_data, content_type=content_type)
        st.write("File ready to be seen in OHIF Viewer.")
    except Exception as e:
        st.error(f"An unexpected error occurred: {e}")

def load_dicom_from_gcs(file_name: str = "dicom_00000001_000.dcm"):
    # Get the blob object
    blob = bucket_load.blob(file_name)

    # Download the file as a bytes object
    dicom_bytes = blob.download_as_bytes()

    # Wrap bytes object into BytesIO (file-like object)
    dicom_stream = io.BytesIO(dicom_bytes)

    # Load the DICOM file 
    ds = pydicom.dcmread(dicom_stream)

    return ds

def png_to_dicom(image_path: str, image_name: str, dicom: str = None):
    if dicom is None:
        ds = load_dicom_from_gcs()
    else:
        ds = load_dicom_from_gcs(dicom)

    jpg_image = Image.open(image_path)  # Open the image using the path
    print("Image Mode:", jpg_image.mode)
    if jpg_image.mode == 'L':
        np_image = np.array(jpg_image.getdata(), dtype=np.uint8)
        ds.Rows = jpg_image.height
        ds.Columns = jpg_image.width
        ds.PhotometricInterpretation = "MONOCHROME1"
        ds.SamplesPerPixel = 1
        ds.BitsStored = 8
        ds.BitsAllocated = 8
        ds.HighBit = 7
        ds.PixelRepresentation = 0
        ds.PixelData = np_image.tobytes()
        ds.save_as(image_name)

    elif jpg_image.mode == 'RGBA':
        np_image = np.array(jpg_image.getdata(), dtype=np.uint8)[:, :3]
        ds.Rows = jpg_image.height
        ds.Columns = jpg_image.width
        ds.PhotometricInterpretation = "RGB"
        ds.SamplesPerPixel = 3
        ds.BitsStored = 8
        ds.BitsAllocated = 8
        ds.HighBit = 7
        ds.PixelRepresentation = 0
        ds.PixelData = np_image.tobytes()
        ds.save_as(image_name)
    elif jpg_image.mode == 'RGB': 
        np_image = np.array(jpg_image.getdata(), dtype=np.uint8)[:, :3] # Remove alpha if present
        ds.Rows = jpg_image.height
        ds.Columns = jpg_image.width
        ds.PhotometricInterpretation = "RGB"
        ds.SamplesPerPixel = 3
        ds.BitsStored = 8
        ds.BitsAllocated = 8
        ds.HighBit = 7
        ds.PixelRepresentation = 0
        ds.PixelData = np_image.tobytes()
        ds.save_as(image_name)
    else:
        raise ValueError("Unsupported image mode:", jpg_image.mode)
    return ds

def save_dicom_to_bytes(dicom):
    dicom_bytes = io.BytesIO()
    dicom.save_as(dicom_bytes)
    dicom_bytes.seek(0)
    return dicom_bytes

def upload_folder_images(original_image_path, enhanced_image_path):
    # Extract the base name of the uploaded image without the extension
    folder_name = os.path.splitext(uploaded_file.name)[0]
    # Create the folder in Cloud Storage
    bucket_result.blob(folder_name + '/').upload_from_string('', content_type='application/x-www-form-urlencoded')
    enhancement_name = enhancement_type.split('_')[-1]
    # Convert images to DICOM
    original_dicom = png_to_dicom(original_image_path, "original_image.dcm")
    enhanced_dicom = png_to_dicom(enhanced_image_path, enhancement_name + ".dcm")

    # Convert DICOM to byte stream for uploading
    original_dicom_bytes = io.BytesIO()
    enhanced_dicom_bytes = io.BytesIO()
    original_dicom.save_as(original_dicom_bytes)
    enhanced_dicom.save_as(enhanced_dicom_bytes)
    original_dicom_bytes.seek(0)
    enhanced_dicom_bytes.seek(0)

    # Upload images to GCS
    upload_to_gcs(original_dicom_bytes, folder_name + '/' + 'original_image.dcm', content_type='application/dicom')
    upload_to_gcs(enhanced_dicom_bytes, folder_name + '/' + enhancement_name + '.dcm', content_type='application/dicom')


def get_mean_std_per_batch(image_path, df, H=320, W=320):
    sample_data = []
    for idx, img in enumerate(df.sample(100)["Image Index"].values):
        # path = image_dir + img
        sample_data.append(
            np.array(keras.utils.load_img(image_path, target_size=(H, W))))

    mean = np.mean(sample_data[0])
    std = np.std(sample_data[0])
    return mean, std

def load_image(img_path, preprocess=True, height=320, width=320):
    mean, std = get_mean_std_per_batch(img_path, df, height, width)
    x = keras.utils.load_img(img_path, target_size=(height, width))
    x = keras.utils.img_to_array(x)
    if preprocess:
        x -= mean
        x /= std
        x = np.expand_dims(x, axis=0)
    return x

def grad_cam(input_model, img_array, cls, layer_name):
    grad_model = tf.keras.models.Model(
        [input_model.inputs],
        [input_model.get_layer(layer_name).output, input_model.output]
    )

    with tf.GradientTape() as tape:
        conv_outputs, predictions = grad_model(img_array)
        loss = predictions[:, cls]

    output = conv_outputs[0]
    grads = tape.gradient(loss, conv_outputs)[0]
    gate_f = tf.cast(output > 0, 'float32')
    gate_r = tf.cast(grads > 0, 'float32')
    guided_grads = gate_f * gate_r * grads

    weights = tf.reduce_mean(guided_grads, axis=(0, 1))

    cam = np.dot(output, weights)

    for index, w in enumerate(weights):
        cam += w * output[:, :, index]

    cam = cv2.resize(cam.numpy(), (320, 320), cv2.INTER_LINEAR)
    cam = np.maximum(cam, 0)
    cam = cam / cam.max()

    return cam


# Compute Grad-CAM
def compute_gradcam(model, img_path, layer_name='bn'):
    preprocessed_input = load_image(img_path)
    predictions = model.predict(preprocessed_input)

    original_image = load_image(img_path, preprocess=False)

    # Assuming you have 14 classes as previously mentioned
    labels = ['Cardiomegaly', 'Emphysema', 'Effusion', 'Hernia', 'Infiltration', 'Mass', 
              'Nodule', 'Atelectasis', 'Pneumothorax', 'Pleural_Thickening', 
              'Pneumonia', 'Fibrosis', 'Edema', 'Consolidation']
    
    for i in range(len(labels)):
        st.write(f"Generating gradcam for class {labels[i]}")
        gradcam = grad_cam(model, preprocessed_input, i, layer_name)
        gradcam = (gradcam * 255).astype(np.uint8)
        gradcam = cv2.applyColorMap(gradcam, cv2.COLORMAP_JET)
        gradcam = cv2.addWeighted(gradcam, 0.5, original_image.squeeze().astype(np.uint8), 0.5, 0)
        st.image(gradcam, caption=f"{labels[i]}: p={predictions[0][i]:.3f}", use_column_width=True)

def calculate_mse(original_image, enhanced_image):
    mse = np.mean((original_image - enhanced_image) ** 2)
    return mse

def calculate_psnr(original_image, enhanced_image):
    mse = calculate_mse(original_image, enhanced_image)
    if mse == 0:
        return float('inf')
    max_pixel_value = 255.0
    psnr = 20 * np.log10(max_pixel_value / np.sqrt(mse))
    return psnr

def calculate_maxerr(original_image, enhanced_image):
    maxerr = np.max((original_image - enhanced_image) ** 2)
    return maxerr

def calculate_l2rat(original_image, enhanced_image):
    l2norm_ratio = np.sum(original_image ** 2) / np.sum((original_image - enhanced_image) ** 2)
    return l2norm_ratio

def process_image(original_image, enhancement_type, fix_monochrome=True):
    if fix_monochrome and original_image.shape[-1] == 3:
        original_image = cv2.cvtColor(original_image, cv2.COLOR_BGR2GRAY)

    image = original_image - np.min(original_image)
    image = image / np.max(original_image)
    image = (image * 255).astype(np.uint8)

    enhanced_image = enhance_image(image, enhancement_type)

    mse = calculate_mse(original_image, enhanced_image)
    psnr = calculate_psnr(original_image, enhanced_image)
    maxerr = calculate_maxerr(original_image, enhanced_image)
    l2rat = calculate_l2rat(original_image, enhanced_image)

    return enhanced_image, mse, psnr, maxerr, l2rat

def apply_clahe(image):
    clahe = cv2.createCLAHE(clipLimit=40.0, tileGridSize=(8, 8))
    return clahe.apply(image)

def invert(image):
    return cv2.bitwise_not(image)

def hp_filter(image, kernel=None):
    if kernel is None:
        kernel = np.array([[-1, -1, -1], [-1, 9, -1], [-1, -1, -1]])
    return cv2.filter2D(image, -1, kernel)

def unsharp_mask(image, radius=5, amount=2):
    def usm(image, radius, amount):
        blurred = cv2.GaussianBlur(image, (0, 0), radius)
        sharpened = cv2.addWeighted(image, 1.0 + amount, blurred, -amount, 0)
        return sharpened
    return usm(image, radius, amount)

def hist_eq(image):
    return cv2.equalizeHist(image)

def enhance_image(image, enhancement_type):
    if enhancement_type == "Invert":
        return invert(image)
    elif enhancement_type == "High Pass Filter":
        return hp_filter(image)
    elif enhancement_type == "Unsharp Masking":
        return unsharp_mask(image)
    elif enhancement_type == "Histogram Equalization":
        return hist_eq(image)
    elif enhancement_type == "CLAHE":
        return apply_clahe(image)
    else:
        raise ValueError(f"Unknown enhancement type: {enhancement_type}")

# Function to add a button to redirect to the URL
def redirect_button(url):
    button = st.button('Go to OHIF Viewer')
    if button:
        st.markdown(f'<meta http-equiv="refresh" content="0;url={url}" />', unsafe_allow_html=True)

def load_model():
    model = tf.keras.models.load_model('./model.h5',custom_objects={'DepthwiseConv2D': tf.keras.layers.DepthwiseConv2D})
    return model

###########################################################################################
########################### Streamlit Interface ###########################################
###########################################################################################


st.sidebar.title("Configuration")
uploaded_file = st.sidebar.file_uploader("Upload Original Image", type=["png", "jpg", "jpeg", "dcm"])
enhancement_type = st.sidebar.selectbox(
    "Enhancement Type", 
    ["Invert", "High Pass Filter", "Unsharp Masking", "Histogram Equalization", "CLAHE"]
)

# File uploader for DICOM files
if uploaded_file is not None:
    if hasattr(uploaded_file, 'name'):
        file_extension = uploaded_file.name.split(".")[-1]  # Get the file extension
        if file_extension.lower() == "dcm":
            # Process DICOM file
            dicom_data = pydicom.dcmread(uploaded_file)
            pixel_array = dicom_data.pixel_array
            # Process the pixel_array further if needed
            # Extract all metadata
            metadata = {elem.keyword: elem.value for elem in dicom_data if elem.keyword}
            metadata_dict = {str(key): str(value) for key, value in metadata.items()}
            df = pd.DataFrame.from_dict(metadata_dict, orient='index', columns=['Value'])

            # Display metadata in the left-most column
            with st.expander("Lihat Metadata"):
                st.write("Metadata:")
                st.dataframe(df)

            # Read the pixel data
            pixel_array = dicom_data.pixel_array
            img_array = pixel_array.astype(float)
            img_array = (np.maximum(img_array, 0) / img_array.max()) * 255.0  # Normalize to 0-255
            img_array = np.uint8(img_array)  # Convert to uint8
            img = Image.fromarray(img_array)

            col1, col2 = st.columns(2)
            # Check the number of dimensions of the image
            if img_array.ndim == 3:
                n_slices = img_array.shape[0]
                if n_slices > 1:
                    slice_ix = st.sidebar.slider('Slice', 0, n_slices - 1, int(n_slices / 2))
                    # Display the selected slice
                    st.image(img_array[slice_ix, :, :], caption=f"Slice {slice_ix}", use_column_width=True)
                else:
                    # If there's only one slice, just display it
                    st.image(img_array[0, :, :], caption="Single Slice Image", use_column_width=True)
            elif img_array.ndim == 2:
                # If the image is 2D, just display it
                with col1:
                    st.image(img_array, caption="Original Image", use_column_width=True)
            else:
                st.error("Unsupported image dimensions")
            
            original_image = img_array

            # Example: convert to grayscale if it's a color image
            if len(pixel_array.shape) > 2:
                pixel_array = pixel_array[:, :, 0]  # Take only the first channel
            # Perform image enhancement and evaluation on pixel_array
            enhanced_image, mse, psnr, maxerr, l2rat = process_image(pixel_array, enhancement_type)
        else:
            # Process regular image file
            original_image = np.array(keras.utils.load_img(uploaded_file, color_mode='rgb' if enhancement_type == "Invert" else 'grayscale'))
            # Perform image enhancement and evaluation on original_image
            enhanced_image, mse, psnr, maxerr, l2rat = process_image(original_image, enhancement_type)
            col1, col2 = st.columns(2)
            with col1:
                st.image(original_image, caption="Original Image", use_column_width=True)
        with col2:
            st.image(enhanced_image, caption='Enhanced Image', use_column_width=True)

        col1, col2 = st.columns(2)
        col3, col4 = st.columns(2)

        col1.metric("MSE", round(mse,3))
        col2.metric("PSNR", round(psnr,3))
        col3.metric("Maxerr", round(maxerr,3))
        col4.metric("L2Rat", round(l2rat,3))

        # Save enhanced image to a file
        enhanced_image_path = "enhanced_image.png"
        cv2.imwrite(enhanced_image_path, enhanced_image)

        
        # Save enhanced image to a file
        enhanced_image_path = "enhanced_image.png"
        cv2.imwrite(enhanced_image_path, enhanced_image)

        # Save original image to a file
        original_image_path = "original_image.png"
        cv2.imwrite(original_image_path, original_image)

    # Add the redirect button
    col1, col2, col3 = st.columns(3)
    with col1:
        redirect_button("https://new-ohif-viewer-k7c3gdlxua-et.a.run.app/")

    with col2:
        if st.button('Auto Detect'):
            name = uploaded_file.name.split("/")[-1].split(".")[0]
            true_bbox_row = df[df['Image Index'] == uploaded_file.name]

            if not true_bbox_row.empty:
                x1, y1 = int(true_bbox_row['Bbox [x']), int(true_bbox_row['y'])
                x2, y2 = int(true_bbox_row['x_max']), int(true_bbox_row['y_max'])
                true_bbox = [x1, y1, x2, y2]
                label = true_bbox_row['Finding Label'].values[0]

                pred_bbox = predict(image)
                iou = cal_iou(true_bbox, pred_bbox)

                image = cv2.rectangle(image, (x1, y1), (x2, y2), (255, 0, 0), 5)  # BLUE
                image = cv2.rectangle(image, (pred_bbox[0], pred_bbox[1]), (pred_bbox[2], pred_bbox[3]), (0, 0, 255), 5)  # RED

                x_pos = int(image.shape[1] * 0.05)
                y_pos = int(image.shape[0] * 0.05)
                font_size = 0.7

                cv2.putText(image, f"IoU: {iou:.4f}", (x_pos, y_pos), cv2.FONT_HERSHEY_SIMPLEX, font_size, (255, 0, 0), 2)
                cv2.putText(image, f"Label: {label}", (x_pos, y_pos + 30), cv2.FONT_HERSHEY_SIMPLEX, font_size, (255, 255, 255), 2)

                st.image(image, channels="BGR")
            else:
                st.write("No bounding box and label found for this image.")
    
    with col3:
        if st.button('Generate Grad-CAM'):
            model = load_model()
            # Compute and show Grad-CAM
            st.write("Generating Grad-CAM visualizations")
            compute_gradcam(model, uploaded_file)