app.py

import gradio as gr
import numpy as np
import cv2
import json
import os
from tensorflow.keras.models import load_model

# ─── Load model ───────────────────────────────────────────────────────────────
model = load_model("custom_cnn.h5")
IMG_SIZE    = 224
NUM_OUTPUTS = model.output_shape[-1]   # auto-detects 3-class or 16-class

# ─── Class / cluster labels ───────────────────────────────────────────────────
# Priority 1: class_labels.json saved alongside the model (from the 16-class notebook)
# Priority 2: fallback cluster names for the 3-class K-Means model
if os.path.exists("class_labels.json"):
    with open("class_labels.json") as f:
        CLASS_NAMES = json.load(f)["classes"]
else:
    # 3-class K-Means cluster model fallback
    CLASS_NAMES = [f"Cluster {i}" for i in range(NUM_OUTPUTS)]

# ─── Which actual pathology classes are dominant in each cluster ──────────────
# These come from analysing your K-Means cluster assignments vs ground-truth labels.
# REPLACE these lists with the real counts from your own cluster analysis notebook.
CLUSTER_DOMINANT = {
    "Cluster 0": [
        ("Normal",                    0.38),
        ("Mild Ventriculomegaly",     0.22),
        ("Arnold–Chiari Malformation",0.15),
        ("Moderate Ventriculomegaly", 0.14),
        ("Hydranencephaly",           0.11),
    ],
    "Cluster 1": [
        ("Severe Ventriculomegaly",   0.35),
        ("Dandy–Walker Malformation", 0.25),
        ("Holoprosencephaly",         0.18),
        ("Agenesis of Corpus Callosum",0.13),
        ("Intracranial Tumors",       0.09),
    ],
    "Cluster 2": [
        ("Intracranial Tumors",       0.30),
        ("Intracranial Hemorrhages",  0.28),
        ("Holoprosencephaly",         0.20),
        ("Dandy–Walker Malformation", 0.12),
        ("Agenesis of Corpus Callosum",0.10),
    ],
}

# For the 16-class model, dominant "classes in cluster" = top-5 softmax outputs
USE_SOFTMAX_DOMINANT = (NUM_OUTPUTS > 3)

# ─── All 16 ground-truth class names for the dropdown ────────────────────────
ALL_GT_CLASSES = [
    "Normal",
    "Mild Ventriculomegaly",
    "Moderate Ventriculomegaly",
    "Severe Ventriculomegaly",
    "Arnold–Chiari Malformation",
    "Hydranencephaly",
    "Agenesis of Corpus Callosum",
    "Dandy–Walker Malformation",
    "Intracranial Tumors",
    "Intracranial Hemorrhages",
    "Holoprosencephaly",
    "Cerebellar Hypoplasia",
    "Microcephaly",
    "Macrocephaly",
    "Lissencephaly",
    "Unknown / Not provided",
]

# ─── Preprocessing — mirrors the paper §3B pipeline ──────────────────────────
def preprocess(image: np.ndarray) -> np.ndarray:
    """Gaussian blur → median filter → CLAHE → normalize [0,1]."""
    if image is None:
        return None
    img = image.astype(np.uint8)
    # To grayscale
    if img.ndim == 3 and img.shape[2] == 3:
        gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
    else:
        gray = img if img.ndim == 2 else img[:, :, 0]
    # §3B-2: Gaussian + median
    blurred  = cv2.GaussianBlur(gray, (5, 5), sigmaX=1.0)
    median   = cv2.medianBlur(blurred, 5)
    # §3B-3: CLAHE
    clahe    = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
    enhanced = clahe.apply(median)
    # Back to RGB float32 [0,1]
    rgb = cv2.cvtColor(enhanced, cv2.COLOR_GRAY2RGB).astype(np.float32) / 255.0
    return rgb

# ─── EMOJI badges for ranks ───────────────────────────────────────────────────
RANK_EMOJI = ["🥇", "🥈", "🥉", "4️⃣", "5️⃣"]

# ─── Progress-bar helper ──────────────────────────────────────────────────────
def pct_bar(value: float, width: int = 28) -> str:
    filled = round(value * width)
    return "█" * filled + "░" * (width - filled)

# ─── Main prediction function ─────────────────────────────────────────────────
def predict(image, actual_class):
    if image is None:
        empty = "Upload an ultrasound image to begin."
        return empty, empty, empty

    # ── Preprocess & predict ──────────────────────────────────────────────────
    proc  = preprocess(image)
    resized = cv2.resize(proc, (IMG_SIZE, IMG_SIZE))
    inp   = np.expand_dims(resized, axis=0)
    probs = model.predict(inp, verbose=0)[0]   # shape: (num_classes,)

    top5_idx = np.argsort(probs)[::-1][:5]
    pred_idx   = top5_idx[0]
    pred_label = CLASS_NAMES[pred_idx]
    confidence = probs[pred_idx] * 100.0

    # ── Panel 1: Prediction cluster ───────────────────────────────────────────
    cluster_lines = [
        "┌─────────────────────────────────────────┐",
        f"│  PREDICTED CLUSTER / CLASS              │",
        "├─────────────────────────────────────────┤",
        f"│  {pred_label:<39} │",
        f"│  Confidence : {confidence:>6.2f}%                    │",
        "└─────────────────────────────────────────┘",
        "",
        "All cluster probabilities:",
        "─" * 43,
    ]
    for i, (cname, p) in enumerate(zip(CLASS_NAMES, probs)):
        marker = " ◀ PREDICTED" if i == pred_idx else ""
        cluster_lines.append(
            f"  {cname:<35}  {p*100:5.1f}%{marker}"
        )
    cluster_text = "\n".join(cluster_lines)

    # ── Panel 2: Top-5 dominant classes ──────────────────────────────────────
    if USE_SOFTMAX_DOMINANT:
        # 16-class model — dominant = top-5 softmax outputs
        dominant = [(CLASS_NAMES[i], float(probs[i])) for i in top5_idx]
        source_note = f"(direct softmax outputs from {NUM_OUTPUTS}-class model)"
    else:
        # 3-class cluster model — look up pre-computed dominant pathologies
        dominant = CLUSTER_DOMINANT.get(
            pred_label,
            [(f"Class {j}", 0.2) for j in range(5)]
        )
        source_note = f"(pathologies most common in {pred_label})"

    top5_lines = [
        f"TOP 5 DOMINANT PATHOLOGY CLASSES  {source_note}",
        "─" * 63,
        "",
    ]
    for rank, (cname, score) in enumerate(dominant):
        bar   = pct_bar(score)
        emoji = RANK_EMOJI[rank]
        top5_lines.append(
            f"  {emoji}  {cname:<40} {bar}  {score*100:5.1f}%"
        )
    top5_text = "\n".join(top5_lines)

    # ── Panel 3: Actual class comparison ─────────────────────────────────────
    if not actual_class or actual_class == "Unknown / Not provided":
        actual_lines = [
            "ℹ️  No ground-truth label provided.",
            "",
            "Select the actual class from the dropdown",
            "on the left to see a correctness check.",
        ]
    else:
        # For cluster model: check if actual class appears in the top-5 dominant list
        dominant_names = [d[0] for d in dominant]
        in_top5 = actual_class in dominant_names

        # For 16-class model: direct label match
        if USE_SOFTMAX_DOMINANT:
            correct = (actual_class == pred_label)
            match_str = "✅  CORRECT PREDICTION" if correct else f"❌  INCORRECT  (model predicted '{pred_label}')"
        else:
            # Cluster model: soft match — is the actual class in the cluster's top-5?
            if in_top5:
                rank_pos = dominant_names.index(actual_class) + 1
                match_str = f"✅  CORRECT CLUSTER  ('{actual_class}' is #{rank_pos} in {pred_label})"
            else:
                match_str = (
                    f"⚠️  PARTIAL MISS  ('{actual_class}' not in top-5 of {pred_label})\n"
                    f"   This may indicate a cluster assignment issue or borderline case."
                )

        actual_lines = [
            "GROUND TRUTH vs PREDICTION",
            "─" * 43,
            "",
            f"  Actual class   :  {actual_class}",
            f"  Predicted      :  {pred_label}  ({confidence:.1f}%)",
            "",
            f"  {match_str}",
            "",
            "─" * 43,
            "Top-5 dominant classes in predicted cluster:",
        ]
        for rank, (cname, score) in enumerate(dominant):
            tick = "  ✓" if cname == actual_class else "   "
            actual_lines.append(f"  {tick} {rank+1}. {cname:<38} {score*100:.1f}%")

    actual_text = "\n".join(actual_lines)

    return cluster_text, top5_text, actual_text


# ─── Gradio UI ────────────────────────────────────────────────────────────────
CSS = """
body, .gradio-container { background: #0d1117 !important; }
.gr-box, .gr-panel    { background: #161b22 !important; border: 1px solid #30363d !important; }
.gr-button            { background: #238636 !important; color: #fff !important; border: none !important; }
.gr-button:hover      { background: #2ea043 !important; }
.output-text textarea { font-family: 'Courier New', monospace !important; font-size: 13px !important;
                        background: #0d1117 !important; color: #e6edf3 !important;
                        border: 1px solid #30363d !important; }
label span            { color: #8b949e !important; }
h1, h2, h3            { color: #e6edf3 !important; }
"""

with gr.Blocks(css=CSS, title="Fetal Brain MRI Classifier 🧠") as demo:
    gr.Markdown("""
    # 🧠 Fetal Brain MRI Classifier
    #### Ultrasound anomaly detection — Standard CNN / Xception transfer learning
    Upload a fetal ultrasound image, optionally select the known ground-truth class, then click **Submit**.
    """)

    with gr.Row():
        # ── Left column: inputs ──────────────────────────────────────────────
        with gr.Column(scale=1):
            image_input = gr.Image(
                type="numpy",
                label="Ultrasound Image",
                image_mode="RGB",
            )
            actual_input = gr.Dropdown(
                choices=ALL_GT_CLASSES,
                value="Unknown / Not provided",
                label="Actual Ground-Truth Class (optional)",
            )
            with gr.Row():
                clear_btn  = gr.Button("Clear")
                submit_btn = gr.Button("Submit", variant="primary")

        # ── Right column: outputs ────────────────────────────────────────────
        with gr.Column(scale=2):
            cluster_out = gr.Textbox(
                label="🏆  Predicted Cluster / Class",
                lines=14,
                interactive=False,
            )
            top5_out = gr.Textbox(
                label="📊  Top 5 Dominant Pathology Classes",
                lines=10,
                interactive=False,
            )
            actual_out = gr.Textbox(
                label="✅  Actual Class Comparison",
                lines=12,
                interactive=False,
            )

    # ── Wire up events ───────────────────────────────────────────────────────
    submit_btn.click(
        fn=predict,
        inputs=[image_input, actual_input],
        outputs=[cluster_out, top5_out, actual_out],
    )
    clear_btn.click(
        fn=lambda: (None, "Unknown / Not provided", "", "", ""),
        inputs=[],
        outputs=[image_input, actual_input, cluster_out, top5_out, actual_out],
    )

demo.launch()