Spaces:
Sleeping
Sleeping
Update streamlit_app.py
Browse files- streamlit_app.py +132 -120
streamlit_app.py
CHANGED
|
@@ -1,140 +1,152 @@
|
|
|
|
|
| 1 |
# streamlit_app.py
|
| 2 |
-
import
|
|
|
|
| 3 |
import numpy as np
|
|
|
|
|
|
|
| 4 |
import tensorflow as tf
|
| 5 |
-
from tensorflow.keras.models import load_model
|
| 6 |
-
import
|
| 7 |
import pydicom
|
| 8 |
-
from
|
|
|
|
| 9 |
|
| 10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 11 |
|
|
|
|
| 12 |
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
# ---------------------------
|
| 16 |
-
@st.cache_resource
|
| 17 |
-
def load_my_model():
|
| 18 |
-
model = load_model("model/best_model.keras", compile=False)
|
| 19 |
-
return model
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
model = load_my_model()
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
# ---------------------------
|
| 26 |
-
# Preprocess image
|
| 27 |
-
# ---------------------------
|
| 28 |
-
def load_image(file):
|
| 29 |
-
"""Loads PNG/JPG/DICOM and returns a grayscale 224x224 normalized array."""
|
| 30 |
-
filename = file.name.lower()
|
| 31 |
-
|
| 32 |
-
if filename.endswith(".dcm"):
|
| 33 |
-
dcm = pydicom.dcmread(file)
|
| 34 |
-
img = dcm.pixel_array.astype(np.float32)
|
| 35 |
-
img = cv2.resize(img, (224, 224))
|
| 36 |
-
img = img / np.max(img)
|
| 37 |
-
return img
|
| 38 |
-
|
| 39 |
-
# PNG / JPG / JPEG
|
| 40 |
-
img = Image.open(file).convert("L")
|
| 41 |
-
img = img.resize((224, 224))
|
| 42 |
-
img = np.array(img).astype(np.float32) / 255.0
|
| 43 |
-
return img
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
# ---------------------------
|
| 47 |
-
# Robust Grad-CAM
|
| 48 |
-
# ---------------------------
|
| 49 |
-
def grad_cam(model, img_array, layer_name=None, eps=1e-8):
|
| 50 |
-
"""
|
| 51 |
-
img_array: (224,224) normalized grayscale → will be expanded to (1,224,224,1)
|
| 52 |
-
"""
|
| 53 |
-
|
| 54 |
-
# Expand dims for model
|
| 55 |
-
x = np.expand_dims(img_array, axis=0) # (1,224,224)
|
| 56 |
-
x = np.expand_dims(x, axis=-1) # (1,224,224,1)
|
| 57 |
-
x = tf.convert_to_tensor(x, dtype=tf.float32)
|
| 58 |
-
|
| 59 |
-
# Auto-detect last conv layer if not provided
|
| 60 |
-
if layer_name is None:
|
| 61 |
-
for layer in reversed(model.layers):
|
| 62 |
-
if hasattr(layer, "output_shape") and len(layer.output_shape) == 4:
|
| 63 |
-
layer_name = layer.name
|
| 64 |
-
break
|
| 65 |
-
|
| 66 |
-
last_conv = model.get_layer(layer_name)
|
| 67 |
-
grad_model = Model([model.inputs], [last_conv.output, model.output])
|
| 68 |
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 73 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 74 |
grads = tape.gradient(loss, conv_outputs)
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
pooled_grads = tf.reduce_mean(grads, axis=(0, 1, 2)).numpy()
|
| 79 |
-
|
| 80 |
-
conv_outputs = conv_outputs[0].numpy() # (H,W,channels)
|
| 81 |
-
|
| 82 |
-
for i in range(len(pooled_grads)):
|
| 83 |
-
conv_outputs[:, :, i] *= pooled_grads[i]
|
| 84 |
-
|
| 85 |
-
cam = np.mean(conv_outputs, axis=-1)
|
| 86 |
cam = np.maximum(cam, 0)
|
| 87 |
-
|
| 88 |
-
cam
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
|
|
|
| 106 |
try:
|
| 107 |
-
|
|
|
|
|
|
|
|
|
|
| 108 |
|
| 109 |
-
|
| 110 |
-
st.image(img, caption="Uploaded Image", use_container_width=True, clamp=True)
|
| 111 |
|
| 112 |
-
|
| 113 |
-
|
| 114 |
|
| 115 |
-
|
| 116 |
-
|
|
|
|
|
|
|
| 117 |
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
st.write(f"**Probability:** {float(pred):.4f}")
|
| 121 |
|
| 122 |
-
|
| 123 |
-
st.subheader("Grad-CAM Heatmap")
|
| 124 |
try:
|
| 125 |
-
cam =
|
| 126 |
-
|
| 127 |
-
np.uint8(255 * cam), cv2.COLORMAP_JET
|
| 128 |
-
)
|
| 129 |
-
heatmap = cv2.cvtColor(heatmap, cv2.COLOR_BGR2RGB)
|
| 130 |
-
|
| 131 |
-
overlay = 0.4 * heatmap + 0.6 * np.stack([img*255]*3, axis=-1)
|
| 132 |
-
overlay = overlay.astype(np.uint8)
|
| 133 |
-
|
| 134 |
-
st.image(overlay, caption="Grad-CAM", use_container_width=True)
|
| 135 |
-
|
| 136 |
except Exception as e:
|
| 137 |
-
st.
|
| 138 |
|
| 139 |
-
|
| 140 |
-
|
|
|
|
| 1 |
+
%%writefile pneumonia_app/streamlit_app.py
|
| 2 |
# streamlit_app.py
|
| 3 |
+
import io
|
| 4 |
+
import os
|
| 5 |
import numpy as np
|
| 6 |
+
from PIL import Image
|
| 7 |
+
import streamlit as st
|
| 8 |
import tensorflow as tf
|
| 9 |
+
from tensorflow.keras.models import load_model
|
| 10 |
+
from tensorflow.keras.applications.densenet import preprocess_input as densenet_preprocess
|
| 11 |
import pydicom
|
| 12 |
+
from pydicom.pixel_data_handlers.util import apply_voi_lut
|
| 13 |
+
import matplotlib.cm as cm
|
| 14 |
|
| 15 |
+
# -------- CONFIG --------
|
| 16 |
+
MODEL_FILENAME = "Model2_exact_serialized.keras" # model file expected in app folder
|
| 17 |
+
IMG_SIZE = (224, 224)
|
| 18 |
+
THRESHOLD = 0.62
|
| 19 |
+
ENABLE_GRADCAM = True
|
| 20 |
+
# ------------------------
|
| 21 |
|
| 22 |
+
st.set_page_config(page_title="Pneumonia Detection (CheXNet)", layout="centered")
|
| 23 |
|
| 24 |
+
st.title("Pneumonia detection (CheXNet)")
|
| 25 |
+
st.write("Upload a chest X-ray (DICOM or PNG/JPG). The app predicts probability of pneumonia.")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
|
| 27 |
+
# ------- utilities -------
|
| 28 |
+
def dicom_to_image_array(dicom_bytes):
|
| 29 |
+
ds = pydicom.dcmread(io.BytesIO(dicom_bytes), force=True)
|
| 30 |
+
try:
|
| 31 |
+
arr = ds.pixel_array
|
| 32 |
+
except Exception as e:
|
| 33 |
+
raise RuntimeError(f"Could not decode DICOM pixel data: {e}")
|
| 34 |
+
if arr.ndim == 3:
|
| 35 |
+
arr = arr[0]
|
| 36 |
+
try:
|
| 37 |
+
arr = apply_voi_lut(arr, ds)
|
| 38 |
+
except Exception:
|
| 39 |
+
pass
|
| 40 |
+
arr = arr.astype(np.float32)
|
| 41 |
+
if getattr(ds, "PhotometricInterpretation", "").upper() == "MONOCHROME1":
|
| 42 |
+
arr = np.max(arr) - arr
|
| 43 |
+
mn, mx = arr.min(), arr.max()
|
| 44 |
+
if mx > mn:
|
| 45 |
+
arr = (arr - mn) / (mx - mn)
|
| 46 |
+
else:
|
| 47 |
+
arr = arr - mn
|
| 48 |
+
arr = (arr * 255.0).clip(0,255).astype(np.uint8)
|
| 49 |
+
return arr
|
| 50 |
+
|
| 51 |
+
def to_rgb_uint8_from_upload(uploaded_file):
|
| 52 |
+
"""Return RGB uint8 (H,W,3) array resized to IMG_SIZE."""
|
| 53 |
+
if uploaded_file is None:
|
| 54 |
+
raise RuntimeError("No file")
|
| 55 |
+
raw = uploaded_file.read()
|
| 56 |
+
# try DICOM
|
| 57 |
+
try:
|
| 58 |
+
ds = pydicom.dcmread(io.BytesIO(raw), stop_before_pixels=True, force=True)
|
| 59 |
+
if hasattr(ds, "PixelData") or getattr(ds, "Rows", None):
|
| 60 |
+
arr = dicom_to_image_array(raw)
|
| 61 |
+
if arr.ndim == 2:
|
| 62 |
+
arr = np.stack([arr]*3, axis=-1)
|
| 63 |
+
pil = Image.fromarray(arr).convert("RGB").resize(IMG_SIZE)
|
| 64 |
+
return np.array(pil)
|
| 65 |
+
except Exception:
|
| 66 |
+
pass
|
| 67 |
+
# fallback normal image
|
| 68 |
+
try:
|
| 69 |
+
pil = Image.open(io.BytesIO(raw)).convert("L").resize(IMG_SIZE)
|
| 70 |
+
arr = np.stack([np.array(pil)]*3, axis=-1)
|
| 71 |
+
return arr.astype(np.uint8)
|
| 72 |
+
except Exception as e:
|
| 73 |
+
raise RuntimeError("Unsupported file format. Upload a DICOM or PNG/JPG.") from e
|
| 74 |
|
| 75 |
+
# -------- model load (cached) --------
|
| 76 |
+
@st.cache_resource
|
| 77 |
+
def load_predict_model(model_path):
|
| 78 |
+
if not os.path.exists(model_path):
|
| 79 |
+
raise FileNotFoundError(f"Model file not found: {model_path}")
|
| 80 |
+
m = load_model(model_path, compile=False)
|
| 81 |
+
return m
|
| 82 |
+
|
| 83 |
+
# Grad-CAM utilities
|
| 84 |
+
def find_last_conv_layer(m):
|
| 85 |
+
for layer in reversed(m.layers):
|
| 86 |
+
out_shape = getattr(layer, "output_shape", None)
|
| 87 |
+
if out_shape and len(out_shape) == 4 and "conv" in layer.name:
|
| 88 |
+
return layer.name
|
| 89 |
+
return m.layers[-3].name
|
| 90 |
+
|
| 91 |
+
def make_gradcam_image(rgb_uint8, model, last_conv_name=None, alpha=0.4, cmap_name="jet"):
|
| 92 |
+
img = rgb_uint8.astype(np.float32)
|
| 93 |
+
if last_conv_name is None:
|
| 94 |
+
last_conv_name = find_last_conv_layer(model)
|
| 95 |
+
grad_model = tf.keras.models.Model([model.inputs], [model.get_layer(last_conv_name).output, model.output])
|
| 96 |
+
x = densenet_preprocess(np.expand_dims(img.astype(np.float32), axis=0))
|
| 97 |
+
with tf.GradientTape() as tape:
|
| 98 |
+
conv_outputs, preds = grad_model(x)
|
| 99 |
+
loss = preds[:, 0]
|
| 100 |
grads = tape.gradient(loss, conv_outputs)
|
| 101 |
+
weights = tf.reduce_mean(grads, axis=(1,2))
|
| 102 |
+
cam = tf.reduce_sum(tf.multiply(weights[:, tf.newaxis, tf.newaxis, :], conv_outputs), axis=-1)
|
| 103 |
+
cam = tf.squeeze(cam).numpy()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 104 |
cam = np.maximum(cam, 0)
|
| 105 |
+
cam_max = cam.max() if cam.max() != 0 else 1e-8
|
| 106 |
+
cam = cam / cam_max
|
| 107 |
+
cam_img = Image.fromarray(np.uint8(cam * 255)).resize((img.shape[1], img.shape[0]), resample=Image.BILINEAR)
|
| 108 |
+
cam_arr = np.array(cam_img).astype(np.float32)/255.0
|
| 109 |
+
colormap = cm.get_cmap(cmap_name)
|
| 110 |
+
heatmap = colormap(cam_arr)[:, :, :3]
|
| 111 |
+
heat_uint8 = np.uint8(heatmap * 255)
|
| 112 |
+
heat_pil = Image.fromarray(heat_uint8).convert("RGBA").resize((img.shape[1], img.shape[0]))
|
| 113 |
+
base_pil = Image.fromarray(np.uint8(img)).convert("RGBA")
|
| 114 |
+
blended = Image.blend(base_pil, heat_pil, alpha=alpha)
|
| 115 |
+
return blended.convert("RGB")
|
| 116 |
+
|
| 117 |
+
# -------- UI elements --------
|
| 118 |
+
col1, col2 = st.columns([1,1])
|
| 119 |
+
with col1:
|
| 120 |
+
uploaded = st.file_uploader("Upload DICOM or PNG/JPG", type=["dcm","png","jpg","jpeg","tif","tiff"])
|
| 121 |
+
with col2:
|
| 122 |
+
thresh = st.number_input("Decision threshold (probability)", min_value=0.0, max_value=1.0, value=float(THRESHOLD), step=0.01)
|
| 123 |
+
|
| 124 |
+
if uploaded is not None:
|
| 125 |
try:
|
| 126 |
+
rgb = to_rgb_uint8_from_upload(uploaded)
|
| 127 |
+
except Exception as e:
|
| 128 |
+
st.error(f"Failed to process file: {e}")
|
| 129 |
+
st.stop()
|
| 130 |
|
| 131 |
+
st.image(rgb, caption="Input (resized)", use_container_width=True)
|
|
|
|
| 132 |
|
| 133 |
+
# load model (cached)
|
| 134 |
+
model = load_predict_model(MODEL_FILENAME)
|
| 135 |
|
| 136 |
+
# predict
|
| 137 |
+
x_pre = densenet_preprocess(np.expand_dims(rgb.astype(np.float32), axis=0))
|
| 138 |
+
prob = float(model.predict(x_pre, verbose=0).ravel()[0])
|
| 139 |
+
pred = int(prob >= thresh)
|
| 140 |
|
| 141 |
+
st.markdown(f"**Pneumonia probability:** `{prob:.4f}`")
|
| 142 |
+
st.markdown(f"**Predicted class (binary):** `{pred}` — **{'Pneumonia' if pred==1 else 'Normal'}**")
|
|
|
|
| 143 |
|
| 144 |
+
if ENABLE_GRADCAM:
|
|
|
|
| 145 |
try:
|
| 146 |
+
cam = make_gradcam_image(rgb, model)
|
| 147 |
+
st.image(cam, caption="Grad-CAM overlay", use_container_width=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 148 |
except Exception as e:
|
| 149 |
+
st.warning(f"Grad-CAM failed: {e}")
|
| 150 |
|
| 151 |
+
else:
|
| 152 |
+
st.info("Upload a DICOM or PNG/JPG image to run inference.")
|