"""
dermascan_app.py
----------------
DermaScan AI — Clinical Skin Lesion Analysis
A feature-rich Streamlit app built on top of the ISIC 2018 U-Net pipeline.
Run:
streamlit run dermascan_app.py
"""
import os
import sys
import io
import warnings
from pathlib import Path
from datetime import datetime
import cv2
import numpy as np
import torch
from PIL import Image
import streamlit as st
import plotly.graph_objects as go
warnings.filterwarnings("ignore")
# ── Path setup ────────────────────────────────────────────────────────────────
PROJECT_ROOT = Path(__file__).parent
sys.path.insert(0, str(PROJECT_ROOT))
from model import UNet
import app_analysis as ana
# ── Inline constants (no config.py needed in Space) ───────────────────────────
MODEL_REPO = "pavanpraneeth/isic-unet"
MODEL_FILE = "best_model.pth"
IMAGE_CHANNELS = 3
MASK_CHANNELS = 1
DEVICE = (
torch.device("cuda") if torch.cuda.is_available()
else torch.device("mps") if torch.backends.mps.is_available()
else torch.device("cpu")
)
# ─────────────────────────────────────────────────────────────────────────────
# Page Config
# ─────────────────────────────────────────────────────────────────────────────
st.set_page_config(
page_title="DermaScan AI",
page_icon="🔬",
layout="wide",
initial_sidebar_state="expanded",
)
# ─────────────────────────────────────────────────────────────────────────────
# Premium CSS
# ─────────────────────────────────────────────────────────────────────────────
st.markdown("""
""", unsafe_allow_html=True)
# ─────────────────────────────────────────────────────────────────────────────
# Helper: small HTML cards
# ─────────────────────────────────────────────────────────────────────────────
def html_card(title: str, content: str) -> str:
return f"""{title}
{content}"""
def pill(text: str, color: str) -> str:
return f'{text}'
def risk_pill(level: str, color: str) -> str:
return f'{level}'
# ─────────────────────────────────────────────────────────────────────────────
# Model Loading (cached)
# ─────────────────────────────────────────────────────────────────────────────
@st.cache_resource(show_spinner="Loading U-Net model…")
def load_model():
from huggingface_hub import hf_hub_download
model = UNet(in_channels=IMAGE_CHANNELS, out_channels=MASK_CHANNELS)
try:
ckpt_path = hf_hub_download(repo_id=MODEL_REPO, filename=MODEL_FILE)
state = torch.load(ckpt_path, map_location=DEVICE)
model.load_state_dict(state["model_state_dict"])
best_dice = state.get("best_val_dice", 0.0)
status = f"Loaded | Val Dice {best_dice:.4f}"
except Exception as e:
status = f"Warning: could not load checkpoint ({e})"
model.eval().to(DEVICE)
return model, DEVICE, status
model, DEVICE, model_status = load_model()
# ─────────────────────────────────────────────────────────────────────────────
# Sidebar
# ─────────────────────────────────────────────────────────────────────────────
with st.sidebar:
# Branding
st.markdown("""
🔬
DermaScan AI
Clinical Skin Lesion Analysis
""", unsafe_allow_html=True)
st.markdown(f"""
""", unsafe_allow_html=True)
st.divider()
# ── Demographics ──
st.markdown("#### 👤 Patient Context *(optional)*")
st.caption("Risk score is adjusted if details are provided.")
age_over_50 = st.checkbox("Age > 50", value=False)
fair_skin = st.checkbox("Fair skin / Fitzpatrick I–II", value=False)
family_history = st.checkbox("Family history of melanoma", value=False)
prev_melanoma = st.checkbox("Previous melanoma", value=False)
high_sun = st.checkbox("High sun exposure history", value=False)
demographics = {
"age_over_50": age_over_50,
"fair_skin": fair_skin,
"family_history": family_history,
"prev_melanoma": prev_melanoma,
"high_sun_exposure":high_sun,
}
st.divider()
# ── Lesion Tracker ──
st.markdown("#### 📅 Evolution Tracker")
st.caption("Upload a previous scan to detect lesion growth.")
prev_upload = st.file_uploader(
"Previous scan (optional)", type=["jpg","jpeg","png"],
key="prev_scan", label_visibility="collapsed"
)
st.divider()
st.markdown("""
⚠️ Disclaimer
DermaScan AI is a computer-aided
screening tool, not a medical
device. Always consult a qualified
dermatologist for clinical evaluation.
""", unsafe_allow_html=True)
# ─────────────────────────────────────────────────────────────────────────────
# Header
# ─────────────────────────────────────────────────────────────────────────────
st.markdown("""
🔬 DermaScan AI
Upload a dermoscopy image for automated ABCDE analysis, clinical measurements,
risk scoring, and explainability heatmaps — powered by a trained U-Net.
""", unsafe_allow_html=True)
# ─────────────────────────────────────────────────────────────────────────────
# Upload Section
# ─────────────────────────────────────────────────────────────────────────────
col_up, col_pre = st.columns([1, 1], gap="large")
with col_up:
uploaded = st.file_uploader(
"Upload dermoscopy image",
type=["jpg","jpeg","png"],
label_visibility="collapsed",
key="main_upload",
)
if uploaded:
pil_img = Image.open(uploaded).convert("RGB")
image_rgb = np.array(pil_img)
st.image(image_rgb, caption="Uploaded Image", use_container_width=True)
with col_pre:
if uploaded:
quality = ana.check_image_quality(image_rgb)
q_color = "#22c55e" if quality["ok"] else ("#f59e0b" if len(quality["issues"]) == 1 else "#ef4444")
q_icon = "✅" if quality["ok"] else ("⚠️" if len(quality["issues"]) == 1 else "❌")
st.markdown(f"""
Image Quality Check
{quality['score']} / 100
{q_icon} {"Good to analyze" if quality['ok'] else "Proceed with caution"}
""", unsafe_allow_html=True)
cq1, cq2, cq3 = st.columns(3)
cq1.metric("Blur Score", quality["blur"])
cq2.metric("Brightness", quality["brightness"])
cq3.metric("Contrast", quality["contrast"])
if quality["issues"]:
for issue in quality["issues"]:
st.warning(issue)
st.markdown("
", unsafe_allow_html=True)
if not uploaded:
st.info("👆 Upload a dermoscopy image to begin analysis.")
st.stop()
# ─────────────────────────────────────────────────────────────────────────────
# Run Analysis
# ─────────────────────────────────────────────────────────────────────────────
run_btn = st.button("🚀 Run DermaScan Analysis", use_container_width=False)
if not run_btn and "results" not in st.session_state:
st.stop()
if run_btn:
with st.spinner("Running segmentation and clinical analysis…"):
# ── Segmentation ──
mask, prob_map = ana.run_segmentation(model, DEVICE, image_rgb)
# ── ABCDE ──
A = ana.abcde_asymmetry(mask)
B = ana.abcde_border(mask)
C = ana.abcde_color(image_rgb, mask)
D = ana.abcde_diameter(mask)
# ── Risk ──
risk = ana.compute_risk(A, B, C, D, demographics)
# ── Measurements ──
meas = ana.compute_measurements(image_rgb, mask)
# ── Visuals ──
overlay = ana.make_overlay(image_rgb, mask)
gradcam = ana.make_gradcam(model, DEVICE, image_rgb)
# ── Evolution (if prev scan provided) ──
evo = None
if prev_upload:
prev_img = np.array(Image.open(prev_upload).convert("RGB"))
prev_mask, _ = ana.run_segmentation(model, DEVICE, prev_img)
evo = ana.compare_masks(prev_mask, mask)
# Store in session
st.session_state["results"] = dict(
mask=mask, prob_map=prob_map, overlay=overlay,
gradcam=gradcam, A=A, B=B, C=C, D=D,
risk=risk, meas=meas, evo=evo, quality=quality,
)
# Pull from session
res = st.session_state["results"]
mask, prob_map = res["mask"], res["prob_map"]
overlay, gradcam = res["overlay"], res["gradcam"]
A, B, C, D = res["A"], res["B"], res["C"], res["D"]
risk, meas, evo = res["risk"], res["meas"], res["evo"]
st.divider()
# ─────────────────────────────────────────────────────────────────────────────
# --- TOP SUMMARY ROW ---
# ─────────────────────────────────────────────────────────────────────────────
s1, s2, s3, s4, s5 = st.columns(5)
s1.metric("Risk Level", risk["level"])
s2.metric("A — Asymmetry", f"{A['score']} / 2", delta=A["label"])
s3.metric("B — Border", f"{B['score']:.2f}", delta=B["label"])
s4.metric("C — Colors", C["count"], delta=C["label"])
s5.metric("Diameter Est.", f"{D['diameter_mm']} mm", delta=D["label"])
st.divider()
# ─────────────────────────────────────────────────────────────────────────────
# --- TABS ---
# ─────────────────────────────────────────────────────────────────────────────
tabs = st.tabs([
"🎯 Overview",
"🔬 ABCDE Analysis",
"📐 Measurements",
"🧠 Explainability",
"📅 Evolution",
"📄 Report",
])
# ══════════════════════════════════════════════════════════════════════════════
# TAB 1 — OVERVIEW
# ══════════════════════════════════════════════════════════════════════════════
with tabs[0]:
ov1, ov2, ov3 = st.columns([1.1, 1.05, 0.95])
with ov1:
st.markdown("**Segmentation Overlay**")
st.image(overlay, use_container_width=True)
st.markdown("**Predicted Mask**")
mask_vis = (mask.astype(np.uint8) * 255)
st.image(mask_vis, use_container_width=True, clamp=True)
with ov2:
gauge_color = risk["color"]
# ── Score number displayed cleanly above gauge ──
st.markdown(f"""
Risk Score
{risk['score']}
/ 10
{risk['level']}
{risk['label']}
""", unsafe_allow_html=True)
# ── Gauge (number hidden — shown above instead) ──
fig = go.Figure(go.Indicator(
mode="gauge",
value=risk["score"],
gauge={
"axis": {
"range": [0, 10],
"tickvals": [0, 2.5, 5, 7.5, 10],
"ticktext": ["0", "2.5", "5", "7.5", "10"],
"tickcolor": "#7a76a8",
"tickfont": {"color": "#7a76a8", "size": 11},
},
"bar": {"color": gauge_color, "thickness": 0.3},
"bgcolor": "#1a1a36",
"borderwidth": 1,
"bordercolor": "rgba(120,100,255,.3)",
"steps": [
{"range": [0, 2.5], "color": "#0d2b12"},
{"range": [2.5, 5.0], "color": "#2b2100"},
{"range": [5.0, 10], "color": "#2b0a0a"},
],
"threshold": {
"line": {"color": "white", "width": 3},
"thickness": 0.85,
"value": risk["score"],
},
},
))
fig.update_layout(
height=200,
margin=dict(l=16, r=16, t=8, b=8),
paper_bgcolor="#09090f",
font={"family": "Inter"},
)
st.plotly_chart(fig, use_container_width=True)
with ov3:
# ABCDE bar chart
criterion_labels = ["A Asymmetry", "B Border", "C Color Var.", "D Diameter"]
raw_vals = [
A["score"] / 2.0,
B["score"],
min(C["count"] / 6.0, 1.0),
float(D["risk"]),
]
bar_colors = [
"#ef4444" if v > 0.5 else "#f59e0b" if v > 0.2 else "#22c55e"
for v in raw_vals
]
fig2 = go.Figure(go.Bar(
x=raw_vals,
y=criterion_labels,
orientation="h",
marker=dict(color=bar_colors, line=dict(color="rgba(0,0,0,0)")),
text=[f"{v:.0%}" for v in raw_vals],
textposition="outside",
textfont=dict(color="#e8e4ff", size=11, family="Inter"),
))
fig2.update_layout(
title=dict(text="ABCD Criteria", font={"color": "#e8e4ff", "size": 13}),
xaxis=dict(range=[0, 1.25], showgrid=False, visible=False),
yaxis=dict(tickfont={"color": "#a89aff", "size": 12}),
plot_bgcolor="#141430",
paper_bgcolor="#09090f",
margin=dict(l=8, r=40, t=36, b=8),
height=210,
)
st.plotly_chart(fig2, use_container_width=True)
if meas:
st.markdown(f"""
Quick Stats
| Area |
{meas['area_mm2']} mm² |
| Perimeter |
{meas['perimeter_mm']} mm |
| Coverage |
{meas['coverage_pct']}% |
""", unsafe_allow_html=True)
# ══════════════════════════════════════════════════════════════════════════════
# TAB 2 — ABCDE ANALYSIS
# ══════════════════════════════════════════════════════════════════════════════
with tabs[1]:
st.markdown("### ABCDE Clinical Criteria")
st.caption("The ABCDE rule is the standard dermatological checklist for melanoma screening. "
"Each criterion is computed directly from the segmentation mask and the original image.")
def abcde_card(letter, name, value_str, label, risk_flag, detail):
icon = "🔴" if risk_flag else "🟢"
color = "#ef4444" if risk_flag else "#22c55e"
st.markdown(f"""
{letter}
{name}
{value_str}
{icon} {label}
{detail}
""", unsafe_allow_html=True)
c1, c2 = st.columns(2)
with c1:
abcde_card(
"A", "Asymmetry",
f"{A['score']} / 2",
A["label"], A["risk"],
f"Horizontal axis asymmetry: {A['axis_h']:.3f} "
f"Vertical axis asymmetry: {A['axis_v']:.3f}
"
"Score of 1 = asymmetric on one axis. Score 2 = both axes. "
"Threshold per axis: 0.18 overlap mismatch ratio."
)
abcde_card(
"C", "Color Variegation",
f"{C['count']} clusters",
C["label"], C["risk"],
f"K-Means clustering (k≤6) applied to lesion pixels. "
f"
Clusters with >3% of lesion pixel population are counted as significant. "
f"≥3 distinct clusters raises concern."
)
with c2:
abcde_card(
"B", "Border Irregularity",
f"{B['score']:.3f}",
B["label"], B["risk"],
f"Irregularity = 1 − (4π·Area / Perimeter²). "
f"Circularity index: {B['circularity']:.3f} "
f"(1.0 = perfect circle). "
"Score > 0.25: irregular. > 0.50: highly irregular."
)
abcde_card(
"D", "Diameter",
f"{D['diameter_mm']} mm",
D["label"], D["risk"],
f"Estimated from pixel area assuming {ana.PIXELS_PER_MM:.0f} px/mm (standard dermoscope). "
f"Area: {D['area_mm2']} mm² ({D['area_px']} px). "
"Clinical threshold: > 6 mm → warning."
)
# Colour palette for C
if C["hex_colors"]:
st.markdown("---")
st.markdown("**Detected Lesion Color Clusters**")
swatches = "".join(
f''
f'{hx}'
for hx in C["hex_colors"]
)
st.markdown(swatches, unsafe_allow_html=True)
# ══════════════════════════════════════════════════════════════════════════════
# TAB 3 — MEASUREMENTS
# ══════════════════════════════════════════════════════════════════════════════
with tabs[2]:
st.markdown("### Clinical Measurements")
st.caption(f"All real-world estimates assume {ana.PIXELS_PER_MM:.0f} px/mm dermoscope calibration.")
if not meas:
st.warning("No lesion detected in the mask — check segmentation threshold.")
st.stop()
m1, m2, m3, m4 = st.columns(4)
m1.metric("Area", f"{meas['area_mm2']} mm²", delta=f"{meas['area_px']} px")
m2.metric("Perimeter", f"{meas['perimeter_mm']} mm")
m3.metric("Image Coverage", f"{meas['coverage_pct']} %")
m4.metric("Mean Lesion Color", meas["mean_color_hex"].upper())
st.markdown("---")
c_det, c_vis = st.columns([1, 1])
with c_det:
st.markdown(f"""
Full Measurement Table
| Area (pixels) |
{meas['area_px']:,} |
| Area (mm²) |
{meas['area_mm2']} |
| Perimeter (px) |
{meas['perimeter_px']} |
| Perimeter (mm) |
{meas['perimeter_mm']} |
| Image Coverage |
{meas['coverage_pct']} % |
| Mean Color (RGB) |
{tuple(meas['mean_color_rgb'])} |
| Color Std Dev |
{tuple(meas['std_color_rgb'])} |
""", unsafe_allow_html=True)
with c_vis:
# Bounding-box visualization
bbox_img = image_rgb.copy()
x, y, bw, bh = meas["bbox"]
cv2.rectangle(bbox_img, (x, y), (x+bw, y+bh), (124, 92, 252), 2)
# Centroid
cx_m = x + bw // 2
cy_m = y + bh // 2
cv2.drawMarker(bbox_img, (cx_m, cy_m), (0, 229, 163), cv2.MARKER_CROSS, 16, 2)
st.image(bbox_img, caption="Bounding box + lesion centroid", use_container_width=True)
# Mean color swatch
mc = meas["mean_color_hex"]
st.markdown(f"""
{mc.upper()}
Mean lesion color
""", unsafe_allow_html=True)
# Probability heatmap
st.markdown("**Segmentation Confidence Map**")
import matplotlib.pyplot as plt
import matplotlib.cm as cm
fig3, ax = plt.subplots(figsize=(5, 4))
fig3.patch.set_facecolor("#09090f")
ax.set_facecolor("#09090f")
im = ax.imshow(prob_map, cmap="magma", vmin=0, vmax=1)
cbar = fig3.colorbar(im, ax=ax)
cbar.ax.yaxis.set_tick_params(color="white")
plt.setp(cbar.ax.yaxis.get_ticklabels(), color="white")
ax.set_title("P(lesion) per pixel", color="#a89aff", fontsize=11)
ax.axis("off")
st.pyplot(fig3, use_container_width=True)
plt.close(fig3)
# ══════════════════════════════════════════════════════════════════════════════
# TAB 4 — EXPLAINABILITY
# ══════════════════════════════════════════════════════════════════════════════
with tabs[3]:
st.markdown("### Model Explainability — Grad-CAM")
st.caption(
"Gradient-weighted Class Activation Mapping (Grad-CAM) reveals which image regions "
"activated the model's last encoder layer most strongly during prediction. "
"Hot regions (red/yellow) had the greatest influence on the segmentation output."
)
e1, e2, e3 = st.columns(3)
with e1:
st.image(image_rgb, caption="Original", use_container_width=True)
with e2:
st.image(gradcam, caption="Grad-CAM Heatmap", use_container_width=True)
with e3:
st.image(overlay, caption="Segmentation", use_container_width=True)
st.markdown("---")
st.markdown(f"""
How to Read This
🔴 Red / Yellow — Regions the encoder focused on most intensely.
🔵 Blue / Dark — Low activation regions, less important to the prediction.
Ideally, the hottest regions should overlap with the visible lesion.
If activation appears in the background, it may indicate the model is relying
on surrounding skin texture — a sign the pretrained encoder would be beneficial.
""", unsafe_allow_html=True)
# ══════════════════════════════════════════════════════════════════════════════
# TAB 5 — EVOLUTION
# ══════════════════════════════════════════════════════════════════════════════
with tabs[4]:
st.markdown("### Lesion Evolution Tracker")
if evo is None:
st.info(
"📅 Upload a **previous scan** in the sidebar to compare this lesion "
"against an earlier timepoint and detect growth."
)
else:
growth_color = "#ef4444" if evo["warning"] else "#22c55e"
ev1, ev2, ev3, ev4 = st.columns(4)
ev1.metric("Previous Area", f"{evo['area_old_mm2']} mm²")
ev2.metric("Current Area", f"{evo['area_new_mm2']} mm²")
ev3.metric("Growth", f"{evo['growth_pct']} %",
delta=("⚠ Growth detected" if evo["warning"] else "Stable"))
ev4.metric("Shape Overlap (IoU)", f"{evo['iou']:.3f}")
if evo["warning"]:
st.error("⚠️ **Significant change detected.** Growth > 20% or shape overlap < 0.70. Please consult a dermatologist.")
else:
st.success("✅ Lesion appears stable compared to the previous scan.")
st.markdown("---")
st.markdown("**Change Map**")
st.image(evo["change_image"], use_container_width=False, width=400,
caption="🟢 Stable | 🔴 New growth | 🔵 Regression")
st.caption("Pixel-level comparison between previous and current segmentation masks.")
# ══════════════════════════════════════════════════════════════════════════════
# TAB 6 — REPORT
# ══════════════════════════════════════════════════════════════════════════════
with tabs[5]:
st.markdown("### Clinical Report")
# Preview
timestamp = datetime.now().strftime("%Y-%m-%d %H:%M")
st.markdown(f"""
Report Preview
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
DERMASCAN AI — CLINICAL LESION REPORT
Generated: {timestamp}
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
IMAGE QUALITY
Score: {res['quality']['score']} / 100 — {"OK" if res['quality']['ok'] else "Issues detected"}
ABCDE ANALYSIS
A Asymmetry: {A['score']} / 2 — {A['label']}
B Border: {B['score']:.3f} — {B['label']}
C Color: {C['count']} clusters — {C['label']}
D Diameter: {D['diameter_mm']} mm — {D['label']}
CLINICAL MEASUREMENTS
Area: {meas.get('area_mm2','—')} mm²
Perimeter: {meas.get('perimeter_mm','—')} mm
Coverage: {meas.get('coverage_pct','—')} %
RISK ASSESSMENT
Score: {risk['score']} / 10.0 — {risk['level']}
{risk['label']}
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
DISCLAIMER: AI screening tool — NOT a medical diagnosis.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
""", unsafe_allow_html=True)
# PDF Download
st.markdown("---")
col_dl1, col_dl2 = st.columns(2)
with col_dl1:
pdf_bytes = ana.generate_pdf(
image_rgb, mask,
res["quality"],
{"A": A, "B": B, "C": C, "D": D},
risk, meas,
)
if pdf_bytes:
st.download_button(
label="⬇️ Download PDF Report",
data=pdf_bytes,
file_name=f"dermascan_report_{datetime.now().strftime('%Y%m%d_%H%M%S')}.pdf",
mime="application/pdf",
use_container_width=True,
)
else:
st.info("Install `fpdf2` for PDF export: `pip install fpdf2`")
with col_dl2:
# Plain-text download (always works)
txt_report = f"""DERMASCAN AI — CLINICAL LESION REPORT
Generated: {timestamp}
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
IMAGE QUALITY
Score: {res['quality']['score']} / 100
Blur: {res['quality']['blur']}
Brightness: {res['quality']['brightness']}
Contrast: {res['quality']['contrast']}
ABCDE ANALYSIS
A Asymmetry: {A['score']} / 2 — {A['label']} (H={A['axis_h']}, V={A['axis_v']})
B Border: {B['score']:.3f} — {B['label']}
C Color: {C['count']} clusters — {C['label']}
D Diameter: {D['diameter_mm']} mm — {D['label']} (Area: {D['area_mm2']} mm²)
CLINICAL MEASUREMENTS
Area: {meas.get('area_mm2','—')} mm² ({meas.get('area_px','—')} px)
Perimeter: {meas.get('perimeter_mm','—')} mm
Coverage: {meas.get('coverage_pct','—')} % of image
Mean Color: {meas.get('mean_color_hex','—').upper()}
RISK ASSESSMENT
Score: {risk['score']} / 10.0
Level: {risk['level']}
Label: {risk['label']}
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
DISCLAIMER: This report is generated by an AI screening tool and does
NOT constitute a medical diagnosis. Always consult a qualified
dermatologist for clinical evaluation.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
"""
st.download_button(
label="⬇️ Download Text Report",
data=txt_report,
file_name=f"dermascan_report_{datetime.now().strftime('%Y%m%d_%H%M%S')}.txt",
mime="text/plain",
use_container_width=True,
)