streamlit_app.py

# streamlit_app.py
import io
import os
import numpy as np
from PIL import Image
import streamlit as st
import tensorflow as tf
from tensorflow.keras.models import load_model
from tensorflow.keras.applications.densenet import preprocess_input as densenet_preprocess
import pydicom
from pydicom.pixel_data_handlers.util import apply_voi_lut
import matplotlib.cm as cm

# -------- CONFIG --------
MODEL_FILENAME = "Model2_exact_serialized.keras"   # model file expected in app folder
IMG_SIZE = (224, 224)
THRESHOLD = 0.62
ENABLE_GRADCAM = True
# ------------------------

st.set_page_config(page_title="Pneumonia Detection (CheXNet)", layout="centered")

st.title("Pneumonia detection (CheXNet)")
st.write("Upload a chest X-ray (DICOM or PNG/JPG). The app predicts probability of pneumonia.")

# ------- utilities -------
def dicom_to_image_array(dicom_bytes):
    ds = pydicom.dcmread(io.BytesIO(dicom_bytes), force=True)
    try:
        arr = ds.pixel_array
    except Exception as e:
        raise RuntimeError(f"Could not decode DICOM pixel data: {e}")
    if arr.ndim == 3:
        arr = arr[0]
    try:
        arr = apply_voi_lut(arr, ds)
    except Exception:
        pass
    arr = arr.astype(np.float32)
    if getattr(ds, "PhotometricInterpretation", "").upper() == "MONOCHROME1":
        arr = np.max(arr) - arr
    mn, mx = arr.min(), arr.max()
    if mx > mn:
        arr = (arr - mn) / (mx - mn)
    else:
        arr = arr - mn
    arr = (arr * 255.0).clip(0,255).astype(np.uint8)
    return arr

def to_rgb_uint8_from_upload(uploaded_file):
    """Return RGB uint8 (H,W,3) array resized to IMG_SIZE."""
    if uploaded_file is None:
        raise RuntimeError("No file")
    raw = uploaded_file.read()
    # try DICOM
    try:
        ds = pydicom.dcmread(io.BytesIO(raw), stop_before_pixels=True, force=True)
        if hasattr(ds, "PixelData") or getattr(ds, "Rows", None):
            arr = dicom_to_image_array(raw)
            if arr.ndim == 2:
                arr = np.stack([arr]*3, axis=-1)
            pil = Image.fromarray(arr).convert("RGB").resize(IMG_SIZE)
            return np.array(pil)
    except Exception:
        pass
    # fallback normal image
    try:
        pil = Image.open(io.BytesIO(raw)).convert("L").resize(IMG_SIZE)
        arr = np.stack([np.array(pil)]*3, axis=-1)
        return arr.astype(np.uint8)
    except Exception as e:
        raise RuntimeError("Unsupported file format. Upload a DICOM or PNG/JPG.") from e

# -------- model load (cached) --------
@st.cache_resource
def load_predict_model(model_path):
    if not os.path.exists(model_path):
        raise FileNotFoundError(f"Model file not found: {model_path}")
    m = load_model(model_path, compile=False)
    return m

# Grad-CAM utilities
def find_last_conv_layer(m):
    for layer in reversed(m.layers):
        out_shape = getattr(layer, "output_shape", None)
        if out_shape and len(out_shape) == 4 and "conv" in layer.name:
            return layer.name
    return m.layers[-3].name

def make_gradcam_image(rgb_uint8, model, last_conv_name=None, alpha=0.4, cmap_name="jet"):
    img = rgb_uint8.astype(np.float32)
    if last_conv_name is None:
        last_conv_name = find_last_conv_layer(model)
    grad_model = tf.keras.models.Model([model.inputs], [model.get_layer(last_conv_name).output, model.output])
    x = densenet_preprocess(np.expand_dims(img.astype(np.float32), axis=0))
    with tf.GradientTape() as tape:
        conv_outputs, preds = grad_model(x)
        loss = preds[:, 0]
    grads = tape.gradient(loss, conv_outputs)
    weights = tf.reduce_mean(grads, axis=(1,2))
    cam = tf.reduce_sum(tf.multiply(weights[:, tf.newaxis, tf.newaxis, :], conv_outputs), axis=-1)
    cam = tf.squeeze(cam).numpy()
    cam = np.maximum(cam, 0)
    cam_max = cam.max() if cam.max() != 0 else 1e-8
    cam = cam / cam_max
    cam_img = Image.fromarray(np.uint8(cam * 255)).resize((img.shape[1], img.shape[0]), resample=Image.BILINEAR)
    cam_arr = np.array(cam_img).astype(np.float32)/255.0
    colormap = cm.get_cmap(cmap_name)
    heatmap = colormap(cam_arr)[:, :, :3]
    heat_uint8 = np.uint8(heatmap * 255)
    heat_pil = Image.fromarray(heat_uint8).convert("RGBA").resize((img.shape[1], img.shape[0]))
    base_pil = Image.fromarray(np.uint8(img)).convert("RGBA")
    blended = Image.blend(base_pil, heat_pil, alpha=alpha)
    return blended.convert("RGB")

# -------- UI elements --------
col1, col2 = st.columns([1,1])
with col1:
    uploaded = st.file_uploader("Upload DICOM or PNG/JPG", type=["dcm","png","jpg","jpeg","tif","tiff"])
with col2:
    thresh = st.number_input("Decision threshold (probability)", min_value=0.0, max_value=1.0, value=float(THRESHOLD), step=0.01)

if uploaded is not None:
    try:
        rgb = to_rgb_uint8_from_upload(uploaded)
    except Exception as e:
        st.error(f"Failed to process file: {e}")
        st.stop()

    st.image(rgb, caption="Input (resized)", use_column_width=False)

    # load model (cached)
    model = load_predict_model(MODEL_FILENAME)

    # predict
    x_pre = densenet_preprocess(np.expand_dims(rgb.astype(np.float32), axis=0))
    prob = float(model.predict(x_pre, verbose=0).ravel()[0])
    pred = int(prob >= thresh)

    st.markdown(f"**Pneumonia probability:** `{prob:.4f}`")
    st.markdown(f"**Predicted class (binary):** `{pred}`  — **{'Pneumonia' if pred==1 else 'Normal'}**")

    if ENABLE_GRADCAM:
        try:
            cam = make_gradcam_image(rgb, model)
            st.image(cam, caption="Grad-CAM overlay", use_column_width=False)
        except Exception as e:
            st.warning(f"Grad-CAM failed: {e}")

else:
    st.info("Upload a DICOM or PNG/JPG image to run inference.")