import io
import os
import requests
import gradio as gr
from PIL import Image
PORT = os.getenv("PORT", "7860")
BACKEND_URL = os.getenv("BACKEND_PREDICT_URL", f"http://127.0.0.1:{PORT}/predict")
def call_backend(image: Image.Image, api_url: str) -> dict:
buf = io.BytesIO()
image.convert("RGB").save(buf, format="PNG")
buf.seek(0)
resp = requests.post(
api_url,
files={"file": ("xray.png", buf.getvalue(), "image/png")},
timeout=60,
)
resp.raise_for_status()
return resp.json()
def render_top_metrics(payload: dict) -> tuple[str, str]:
model_arm = payload.get("selected_model", "Unknown")
latency_ms = max(payload.get("latency_ms", 0), 0)
latency_html = f'{latency_ms} ms
'
if latency_ms < 0:
latency_html = f'''
{latency_ms} ms
Bug: negative latency
Timing value is invalid — likely a clock sync issue.
Fix the timer before showing this to patients.
'''
arm_html = f'''
MODEL (A/B ARM)
{model_arm}
'''
lat_html = f'''
INFERENCE LATENCY
{latency_html}
'''
return arm_html, lat_html
def render_quality(payload: dict) -> str:
drift = payload.get("drift", {})
alert = drift.get("drift_alert", "NORMAL")
if alert == "DRIFT_DETECTED":
return '''
INPUT IMAGE QUALITY CHECK
⚠ Out of Distribution
Distribution differs from training baseline
What this means for the patient
The model has never seen images like this. Predictions may be unreliable. Proceed with extreme caution.
'''
return '''
INPUT IMAGE QUALITY CHECK
✓ Acceptable
Distribution matches training baseline
What this means for the patient
This does not mean your lungs are normal. It only means the uploaded X-ray image is technically readable by the model — the contrast, resolution, and framing look similar to images it was trained on.
'''
_TIER_COLORS = {
"Pneumonia": "#ef4444", "Pneumothorax": "#ef4444", "Mass": "#ef4444", "Nodule": "#ef4444",
"Effusion": "#f59e0b", "Cardiomegaly": "#f59e0b", "Consolidation": "#f59e0b",
}
_TIER_DESC = {
"Pneumonia": "Infection inflaming air sacs, potentially filling with fluid.",
"Pneumothorax": "Collapsed lung — air leaked into the space between lung and chest wall.",
"Mass": "Larger opacity (>3cm). Not actionable at this score.",
"Nodule": "A small round opacity — may be benign or require follow-up.",
"Effusion": "Fluid build-up in the pleural space.",
"Cardiomegaly": "Enlarged heart — could indicate underlying heart conditions.",
"Consolidation": "Region of normally compressible lung tissue filled with liquid.",
"Infiltration": "Fluid or tissue density where there should be air — could suggest infection, inflammation, or early pneumonia.",
}
def render_top3(payload: dict) -> str:
top3 = payload.get("top_predictions", [])
inconclusive = payload.get("inconclusive_scan", False)
html = ''
html += '
'
html += '
TOP-3 PREDICTED PATHOLOGIES
'
html += '
● Critical ● Significant ● Standard
'
html += '
'
for p in top3:
lbl = p["label"]
conf = p["confidence"]
col = _TIER_COLORS.get(lbl, "#3b82f6")
desc = _TIER_DESC.get(lbl, "A generic chest anomaly flagged by the model.")
conf_text = f"Confidence is very low ({conf}%)." if conf < 10 else f"Confidence: {conf}%."
flagged = '
Flagged finding' if p["detected"] else ''
html += f'''
● {lbl} {flagged}
{conf}%
{desc} {conf_text}
'''
if inconclusive:
html += '''
All confidence scores are below 10% — interpret with extreme caution
Scores this low indicate the model is uncertain. A well-calibrated model should ideally show >50% before a finding is clinically noteworthy. These results should not drive any clinical decision without radiologist review.
'''
html += '
'
html += '
ALL 14 SCREENED CONDITIONS
'
html += '
'
for p in all_preds:
lbl = p["label"]
conf = p["confidence"]
det = p["detected"]
if det:
bg, color = "#78350f", "#fcd34d" # Yellow highlight
else:
bg, color = "#374151", "#9ca3af" # Gray
text = f"{lbl} {conf}%" if det else lbl
html += f'
{text}
'
html += '
'
html += '
Grey chips = below detection threshold. Not shown does not mean absent.
'
html += '
MODEL CALIBRATION & BENCHMARKING
Model calibrated on NIH ChestX-ray14. Historical AUC benchmarks:
- Cardiomegaly: ~0.81 AUC (Highly reliable)
- Infiltration: ~0.70 AUC (Historically lower reliability)
- Pneumonia: ~0.68 AUC
Note for Clinicians: Predictions for "Infiltration" and "Pneumonia" carry higher uncertainty due to lower historical AUC. Radiologist verification is mandatory.
'''
def predict(image: Image.Image, api_url: str):
if image is None:
raise gr.Error("Please upload a chest X-ray image.")
try:
payload = call_backend(image, api_url)
except Exception as exc:
raise gr.Error(f"Backend error: {str(exc)}") from exc
arm_html, lat_html = render_top_metrics(payload)
qc_html = render_quality(payload)
top3_html = render_top3(payload)
classes_html = render_14_classes(payload)
calib_html = render_calibration_info()
ethics_html = """
ETHICAL & CLINICAL COMPLIANCE
Data Privacy: HIPAA-compliant processing. No PHI is stored.
Model Bias: Regular audits for demographic parity across age and gender.
Clinical Intent: Screening assistant only. Not for primary diagnosis.
"""
summary_html = f'CLINICAL SUMMARY
{payload.get("summary", "Screening complete.")}
⚖️ Responsible AI notice
This system assists in prediction, it does not confirm diagnoses. Results are statistical estimates and must be reviewed by a qualified radiologist or clinician before any medical decision is made. Do not use these results to self-diagnose.
''')
submit_btn.click(
fn=predict,
inputs=[image_input, api_url],
outputs=[arm_out, lat_out, summary_out, ethics_out, qc_out, top3_out, classes_out, calib_out],
)
if __name__ == "__main__":
port = int(os.getenv("GRADIO_PORT", "7860"))
demo.launch(server_name="0.0.0.0", server_port=port)