app.py

import os
import csv
import zipfile
import subprocess
import uuid
import shutil
from pathlib import Path
from typing import Tuple, Dict, List, Optional

import gradio as gr
import SimpleITK as sitk
from huggingface_hub import hf_hub_download

import spaces


# Dummy function to satisfy HF Spaces GPU detection during startup
@spaces.GPU
def _init_gpu():
    """Dummy function to ensure Spaces detects GPU usage at startup."""
    import torch
    return torch.cuda.is_available() if torch.cuda.is_available() else True


# =========================
# App config
# =========================
APP_NAME = "Multiple Sclerosis Lesion Tracker"
ALLOW_CPU_FALLBACK = True   # If CUDA fails for FLAMeS, fall back to CPU (slower)

# Per-job subfolders
STRIPPED_DIR   = "stripped"
REGISTERED_DIR = "registered"
SEG_DIR        = "seg_flames"
DIFF_SEG_DIR   = "seg_flames/diff_seg"

# Persistent roots (override via env vars)
JOBS_ROOT    = Path(os.getenv("MSAPP_JOBS_ROOT", Path.cwd() / "jobs")).resolve()
MODEL_ROOT   = Path(os.getenv("MSAPP_MODEL_ROOT", Path.home() / ".cache" / "msapp" / "flames_model")).resolve()
BIN_ROOT     = Path(os.getenv("MSAPP_BIN_DIR",   Path.home() / ".cache" / "msapp" / "bin")).resolve()

# Hugging Face artifact
HF_REPO      = "FrancescoLR/FLAMeS-model"
HF_ZIP_NAME  = "Dataset004_WML.zip"
NNUNET_DS_ID = "004"          # nnUNet dataset id expected by the model
NNUNET_CFG   = "3d_fullres"
NNUNET_TR    = "nnUNetTrainer_8000epochs"

# SynthStrip model (optional for Nipype; REQUIRED for some CLI builds)
SYNTHSTRIP_MODEL = Path(
    os.getenv("MSAPP_SYNTHSTRIP_MODEL", "").strip() or ""
).expanduser() if os.getenv("MSAPP_SYNTHSTRIP_MODEL") else Path("")


# =========================
# Utilities
# =========================
def ensure_dirs() -> None:
    JOBS_ROOT.mkdir(parents=True, exist_ok=True)
    MODEL_ROOT.mkdir(parents=True, exist_ok=True)
    BIN_ROOT.mkdir(parents=True, exist_ok=True)

# After ensure_dirs(), add:
try:
    _init_gpu()
except Exception:
    pass


def is_zip(path: Path) -> bool:
    return path.suffix.lower() == ".zip"


def get_dcm2niix_bin() -> Optional[str]:
    p = os.getenv("DCM2NIIX_BIN")
    if p and Path(p).exists():
        return str(Path(p).resolve())
    return shutil.which("dcm2niix")


# ---- SynthStrip (NiPreps) without Docker ----
def ensure_synthstrip_available() -> None:
    """
    Ensure SynthStrip from NiPreps is available, either:
      - Python interface (preferred):  pip install 'nipreps-synthstrip[nipype]'
      - CLI executable fallback:       pip install nipreps-synthstrip
    """
    try:
        import importlib  # noqa: F401
        importlib.import_module("nipreps.synthstrip.wrappers.nipype")
        return
    except Exception:
        pass

    if shutil.which("nipreps-synthstrip") is None:
        raise RuntimeError(
            "SynthStrip is not available. Install one of:\n"
            "  • Python/Nipype:  pip install 'nipreps-synthstrip[nipype]'\n"
            "  • CLI only:       pip install nipreps-synthstrip\n"
            "and ensure 'nipreps-synthstrip' is on PATH."
        )


def skull_strip_synthstrip(in_nii: Path, out_dir: Path, prefix: str) -> Tuple[Path, Path]:
    """
    Run SynthStrip via NiPreps (Nipype interface if available, else CLI).
    Honors SYNTHSTRIP_MODEL if set.
    Returns (stripped_path, mask_path).
    """
    out_dir.mkdir(parents=True, exist_ok=True)
    stripped = out_dir / f"{prefix}_stripped.nii.gz"
    mask     = out_dir / f"{prefix}_mask.nii.gz"

    model_path = None
    if SYNTHSTRIP_MODEL and str(SYNTHSTRIP_MODEL).strip():
        if not SYNTHSTRIP_MODEL.exists():
            raise RuntimeError(f"SynthStrip model not found at: {SYNTHSTRIP_MODEL}")
        model_path = str(SYNTHSTRIP_MODEL)

    # Prefer Python/Nipype interface
    try:
        from nipreps.synthstrip.wrappers.nipype import SynthStrip  # type: ignore
        kwargs = dict(in_file=str(in_nii), out_file=str(stripped), out_mask=str(mask))
        # Use model if the interface supports it
        if model_path:
            try:
                kwargs["model"] = model_path
            except Exception:
                pass
        # Older/newer APIs may or may not support no_csf; try but don't fail if absent
        try:
            kwargs["no_csf"] = True
        except Exception:
            pass

        node = SynthStrip(**kwargs)
        res = node.run()  # executes interface

        # Ensure outputs exist (some versions may write to alternate paths)
        if not stripped.exists():
            out_file = getattr(res.outputs, "out_file", None)
            if out_file and Path(out_file).exists():
                shutil.copy2(out_file, stripped)
        if not mask.exists():
            out_mask = getattr(res.outputs, "out_mask", None)
            if out_mask and Path(out_mask).exists():
                shutil.copy2(out_mask, mask)
        if not stripped.exists() or not mask.exists():
            raise RuntimeError("SynthStrip (nipype) finished but outputs not found.")
        return stripped, mask
    except Exception:
        # Fall back to CLI
        pass

    # ---- CLI fallback ----
    # Your CLI requires a model; pass it or fail fast.
    if not model_path:
        raise RuntimeError(
            "Your SynthStrip CLI requires a model. Set MSAPP_SYNTHSTRIP_MODEL to the .pt file, e.g.\n"
            "export MSAPP_SYNTHSTRIP_MODEL='/home/karlo/Desktop/app_flames_10/synthstrip_model/synthstrip.nocsf.1.pt'"
        )

    cmd = [
        shutil.which("nipreps-synthstrip") or "nipreps-synthstrip",
        "-i", str(in_nii),
        "-o", str(stripped),
        "-m", str(mask),          # CLI expects -m for mask
        "--model", model_path,    # provide your local model file
        # Optional: threads/border if desired:
        # "-n", str(os.cpu_count() or 1),
        # "-b", "0",
    ]
    run = subprocess.run(cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True)
    if run.returncode != 0:
        raise RuntimeError(f"SynthStrip (CLI) failed:\n{run.stdout}")
    if not stripped.exists() or not mask.exists():
        raise RuntimeError("SynthStrip (CLI) finished but outputs not found.")
    return stripped, mask


def ensure_flames_model() -> None:
    """
    Ensure FLAMeS weights (Dataset004_WML) exist under MODEL_ROOT.
    nnUNet expects: MODEL_ROOT / 'Dataset004_WML' / ...
    """
    ensure_dirs()
    ds_dir = MODEL_ROOT / "Dataset004_WML"
    if ds_dir.exists():
        return
    zip_path = hf_hub_download(repo_id=HF_REPO, filename=HF_ZIP_NAME, cache_dir=str(MODEL_ROOT))
    subprocess.run(["unzip", "-o", zip_path, "-d", str(MODEL_ROOT)], check=True)


def stage_upload(src_path: Path, incoming_dir: Path, name_hint: str) -> Path:
    """
    Copy an uploaded file (which may live under /tmp/gradio) into the job's incoming dir.
    Returns the staged local path (under JOBS_ROOT).
    """
    incoming_dir.mkdir(parents=True, exist_ok=True)
    dst = incoming_dir / f"{name_hint}{src_path.suffix}"
    # If double suffix (.nii.gz) keep it
    if "".join(src_path.suffixes[-2:]).lower() == ".nii.gz":
        dst = incoming_dir / f"{name_hint}.nii.gz"
    shutil.copy2(src_path, dst)
    return dst


def convert_dicom_zip_to_nifti(zip_path: Path, out_dir: Path) -> Path:
    extract_dir = out_dir / "dicom_extracted"
    extract_dir.mkdir(parents=True, exist_ok=True)
    with zipfile.ZipFile(zip_path, "r") as zf:
        zf.extractall(extract_dir)

    nifti_dir = out_dir / "nifti"
    nifti_dir.mkdir(parents=True, exist_ok=True)

    dcm2 = get_dcm2niix_bin()
    if not dcm2:
        raise RuntimeError(
            "dcm2niix not found. Set DCM2NIIX_BIN to its absolute path "
            "(e.g., /home/karlo/anaconda3/envs/nnunet-env/bin/dcm2niix) "
            "or launch the app from the activated env."
        )

    run = subprocess.run(
        [dcm2, "-d", "9", "-z", "y", "-f", "series_%p_%s", "-o", str(nifti_dir), str(extract_dir)],
        stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True
    )
    if run.returncode != 0:
        raise RuntimeError(f"dcm2niix failed:\n{run.stdout}")

    candidates = sorted(nifti_dir.glob("*.nii.gz"), key=lambda p: p.stat().st_size, reverse=True)
    if not candidates:
        raise RuntimeError("No NIfTI produced by dcm2niix.")
    return candidates[0]


def read_float(path: Path) -> sitk.Image:
    return sitk.Cast(sitk.ReadImage(str(path)), sitk.sitkFloat32)


def register_rigid_affine(prev_stripped: Path, new_stripped: Path, reg_dir: Path) -> Tuple[Path, sitk.Transform, sitk.Image]:
    """Register new->prev (rigid then affine). Returns (registered_path, affine_tx, fixed_img)."""
    reg_dir.mkdir(parents=True, exist_ok=True)
    fixed = read_float(prev_stripped)
    moving = read_float(new_stripped)

    # Rigid
    initial = sitk.CenteredTransformInitializer(
        fixed, moving, sitk.VersorRigid3DTransform(),
        sitk.CenteredTransformInitializerFilter.GEOMETRY
    )
    R = sitk.ImageRegistrationMethod()
    R.SetMetricAsMattesMutualInformation(50)
    R.SetMetricSamplingStrategy(R.RANDOM)
    R.SetMetricSamplingPercentage(0.1)
    R.SetInterpolator(sitk.sitkLinear)
    R.SetOptimizerAsRegularStepGradientDescent(learningRate=2.0, minStep=1e-4, numberOfIterations=200, relaxationFactor=0.5)
    R.SetOptimizerScalesFromPhysicalShift()
    R.SetShrinkFactorsPerLevel([4, 2, 1])
    R.SetSmoothingSigmasPerLevel([2, 1, 0]); R.SmoothingSigmasAreSpecifiedInPhysicalUnitsOn()
    R.SetInitialTransform(initial, inPlace=False)
    rigid_tx = R.Execute(fixed, moving)

    if rigid_tx.GetName() == "CompositeTransform":
        rigid_only = rigid_tx.GetNthTransform(0)
    else:
        rigid_only = rigid_tx
    if rigid_only.GetName() != "VersorRigid3DTransform":
        rigid_only = sitk.VersorRigid3DTransform(rigid_only)

    # Affine (seed from rigid)
    A = sitk.ImageRegistrationMethod()
    A.SetMetricAsMattesMutualInformation(50)
    A.SetMetricSamplingStrategy(A.RANDOM)
    A.SetMetricSamplingPercentage(0.1)
    A.SetInterpolator(sitk.sitkLinear)
    A.SetOptimizerAsRegularStepGradientDescent(learningRate=1.0, minStep=1e-4, numberOfIterations=150, relaxationFactor=0.5)
    A.SetOptimizerScalesFromPhysicalShift()
    A.SetShrinkFactorsPerLevel([4, 2, 1])
    A.SetSmoothingSigmasPerLevel([2, 1, 0]); A.SmoothingSigmasAreSpecifiedInPhysicalUnitsOn()
    affine_init = sitk.AffineTransform(3)
    affine_init.SetMatrix(rigid_only.GetMatrix())
    affine_init.SetTranslation(rigid_only.GetTranslation())
    affine_init.SetCenter(rigid_only.GetCenter())
    A.SetInitialTransform(affine_init, inPlace=False)
    affine_tx = A.Execute(fixed, moving)

    # Resample moving into fixed space
    registered_path = reg_dir / "new_in_prev_space.nii.gz"
    resampler = sitk.ResampleImageFilter()
    resampler.SetReferenceImage(fixed)
    resampler.SetInterpolator(sitk.sitkLinear)
    resampler.SetTransform(affine_tx)
    registered_img = resampler.Execute(moving)
    sitk.WriteImage(registered_img, str(registered_path))
    return registered_path, affine_tx, fixed


@spaces.GPU(duration=300)
def run_flames_single(input_nii: Path, out_mask_path: Path, device: str = "cuda") -> Path:
    """Run FLAMeS (nnUNetv2) on a single input NIfTI and write a mask. Uses shared MODEL_ROOT."""
    with (Path(input_nii).open("rb")):
        pass
    import tempfile
    with tempfile.TemporaryDirectory() as in_dir, tempfile.TemporaryDirectory() as out_dir:
        in_img = Path(in_dir) / "image_0000.nii.gz"
        shutil.copy2(input_nii, in_img)
        env = os.environ.copy()
        env["nnUNet_results"] = str(MODEL_ROOT)  # shared cache with weights
        cmd = [
            "nnUNetv2_predict",
            "-i", in_dir, "-o", out_dir,
            "-d", NNUNET_DS_ID,
            "-c", NNUNET_CFG,
            "-tr", NNUNET_TR,
            "-device", device
        ]
        run = subprocess.run(cmd, env=env, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True)
        if run.returncode != 0 and device == "cuda" and ALLOW_CPU_FALLBACK:
            cmd[-1] = "cpu"
            run = subprocess.run(cmd, env=env, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True)
        if run.returncode != 0:
            raise RuntimeError(f"nnUNetv2_predict failed:\n{run.stdout}")
        pred = Path(out_dir) / "image.nii.gz"
        if not pred.exists():
            raise RuntimeError("nnUNet did not produce image.nii.gz")
        out_mask_path.parent.mkdir(parents=True, exist_ok=True)
        shutil.move(pred, out_mask_path)
        return out_mask_path


def build_diff_maps(prev_mask: Path, new_mask_in_prev_space: Path, out_root: Path,
                    dilate_prev_radius_vox: int = 1,           # used for NEW only
                    min_lesion_vol_ml: float = 0.01,
                    edge_buffer_vox_for_resolved: int = 1,     # suppress rim artifacts
                    use_26_connectivity: bool = False):
    out_dir = out_root / DIFF_SEG_DIR
    out_dir.mkdir(parents=True, exist_ok=True)

    prev_img = sitk.ReadImage(str(prev_mask))
    new_img  = sitk.ReadImage(str(new_mask_in_prev_space))
    prev_bin = sitk.Cast(sitk.BinaryThreshold(prev_img, 1, 1_000_000, 1, 0), sitk.sitkUInt8)
    new_bin  = sitk.Cast(sitk.BinaryThreshold(new_img , 1, 1_000_000, 1, 0), sitk.sitkUInt8)

    # Match geometry if needed
    def geom_tuple(img): return (img.GetSize(), img.GetSpacing(), img.GetOrigin(), tuple(img.GetDirection()))
    if geom_tuple(prev_bin) != geom_tuple(new_bin):
        new_bin = sitk.Resample(new_bin, prev_bin, sitk.Transform(3, sitk.sitkIdentity),
                                sitk.sitkNearestNeighbor, 0, sitk.sitkUInt8)

    # Conservative NEW: dilate baseline (helps avoid false-new)
    P_for_new = prev_bin
    if dilate_prev_radius_vox > 0:
        P_for_new = sitk.BinaryDilate(P_for_new, [dilate_prev_radius_vox]*3)

    # RESOLVED: do NOT dilate baseline; instead buffer away the new mask
    N_buffered = new_bin
    if edge_buffer_vox_for_resolved > 0:
        N_buffered = sitk.BinaryDilate(new_bin, [edge_buffer_vox_for_resolved]*3)

    # Sets
    stable      = sitk.And(prev_bin, new_bin)
    new_raw     = sitk.And(new_bin,  sitk.BinaryNot(P_for_new))
    resolved_raw= sitk.And(prev_bin, sitk.BinaryNot(N_buffered))
    xor_mask    = sitk.Xor(prev_bin, new_bin)

    # Voxel volume (mL)
    sx, sy, sz = prev_bin.GetSpacing()
    voxel_ml = (sx * sy * sz) / 1000.0

    # Connected components helper using the filter API (version-safe)
    def connected_components(bin_img, fully_connected):
        cc_filter = sitk.ConnectedComponentImageFilter()
        if fully_connected:
            cc_filter.FullyConnectedOn()
        else:
            cc_filter.FullyConnectedOff()
        return cc_filter.Execute(bin_img)

    def min_volume_filter(bin_img):
        if min_lesion_vol_ml <= 0:
            return sitk.Cast(bin_img > 0, sitk.sitkUInt8)
        cc = connected_components(bin_img, fully_connected=use_26_connectivity)
        stats = sitk.LabelShapeStatisticsImageFilter(); stats.Execute(cc)
        keep = sitk.Image(cc.GetSize(), sitk.sitkUInt8); keep.CopyInformation(cc)
        for lbl in stats.GetLabels():
            if stats.GetNumberOfPixels(lbl) * voxel_ml >= min_lesion_vol_ml:
                keep = sitk.Or(keep, sitk.Equal(cc, lbl))
        return sitk.Cast(keep > 0, sitk.sitkUInt8)

    new_f      = min_volume_filter(new_raw)
    resolved_f = min_volume_filter(resolved_raw)

    paths = {
        "new_only":      out_dir / "new_lesions_mask.nii.gz",
        "resolved_only": out_dir / "resolved_lesions_mask.nii.gz",
        "xor":           out_dir / "xor_diff_mask.nii.gz",
        "stable":        out_dir / "stable_overlap_mask.nii.gz",
        "combined":      out_dir / "combined_label_map.nii.gz",
        "new_csv":       out_dir / "new_lesions_summary.csv",
        "resolved_csv":  out_dir / "resolved_lesions_summary.csv",
    }

    sitk.WriteImage(new_f,      str(paths["new_only"]))
    sitk.WriteImage(resolved_f, str(paths["resolved_only"]))
    sitk.WriteImage(xor_mask,   str(paths["xor"]))
    sitk.WriteImage(stable,     str(paths["stable"]))

    combined = sitk.Image(prev_bin.GetSize(), sitk.sitkUInt8); combined.CopyInformation(prev_bin)
    combined = sitk.Add(combined, sitk.Multiply(stable,     1))
    combined = sitk.Add(combined, sitk.Multiply(new_f,      2))
    combined = sitk.Add(combined, sitk.Multiply(resolved_f, 3))
    sitk.WriteImage(combined, str(paths["combined"]))

    # CSV writers
    def write_csv(bin_img, csv_path: Path):
        cc = connected_components(bin_img, fully_connected=use_26_connectivity)
        rl = sitk.RelabelComponent(cc, sortByObjectSize=True)
        stats = sitk.LabelShapeStatisticsImageFilter(); stats.Execute(rl)
        with open(csv_path, "w", newline="") as f:
            w = csv.writer(f)
            w.writerow(["lesion_id","voxel_count","volume_mL","centroid_x_mm","centroid_y_mm","centroid_z_mm"])
            for lbl in stats.GetLabels():
                vox = stats.GetNumberOfPixels(lbl); vol_ml = vox * voxel_ml
                cx, cy, cz = stats.GetCentroid(lbl)
                w.writerow([int(lbl), int(vox), float(vol_ml), float(cx), float(cy), float(cz)])

    write_csv(new_f,      paths["new_csv"])
    write_csv(resolved_f, paths["resolved_csv"])
    return paths


# ---- helper to count lesions in a mask ----
def count_lesions_in_mask(mask_path: Path,
                          min_lesion_vol_ml: float = 0.0,
                          use_26_connectivity: bool = False) -> int:
    """
    Count connected components in a binary lesion mask (values > 0).
    Applies a minimum lesion volume filter if > 0.
    """
    img = sitk.ReadImage(str(mask_path))
    bin_img = sitk.Cast(sitk.BinaryThreshold(img, 1, 1_000_000, 1, 0), sitk.sitkUInt8)

    # Voxel volume (mL)
    sx, sy, sz = bin_img.GetSpacing()
    voxel_ml = (sx * sy * sz) / 1000.0

    cc_filter = sitk.ConnectedComponentImageFilter()
    if use_26_connectivity:
        cc_filter.FullyConnectedOn()
    else:
        cc_filter.FullyConnectedOff()
    cc = cc_filter.Execute(bin_img)

    stats = sitk.LabelShapeStatisticsImageFilter()
    stats.Execute(cc)

    count = 0
    for lbl in stats.GetLabels():
        vox = stats.GetNumberOfPixels(lbl)
        vol_ml = vox * voxel_ml
        if vol_ml >= min_lesion_vol_ml:
            count += 1
    return count


def package_selected(job_dir: Path,
                     prev_stripped: Path,
                     registered_new: Path,
                     diff_paths: Dict[str, Path],
                     prev_mask_flames: Optional[Path] = None,
                     new_mask_flames: Optional[Path] = None,
                     zip_name: str = "outputs.zip") -> Path:
    """Zip key deliverables (optionally include FLAMeS masks)."""
    zpath = job_dir / zip_name
    to_add = [
        prev_stripped,
        registered_new,
        diff_paths["new_only"],
        diff_paths["resolved_only"],
        diff_paths["xor"],
        diff_paths["stable"],
        diff_paths["combined"],
    ]

    if prev_mask_flames is not None:
        to_add.append(prev_mask_flames)
    if new_mask_flames is not None:
        to_add.append(new_mask_flames)

    with zipfile.ZipFile(zpath, "w", zipfile.ZIP_DEFLATED) as zf:
        for p in to_add:
            if not p.exists():
                raise RuntimeError(f"Expected output missing: {p}")
            zf.write(p, p.relative_to(job_dir))
    return zpath


# =========================
# Main pipeline (UI callback)
# =========================
def _redact_paths(s: str) -> str:
    """Hide absolute paths from user-visible messages."""
    for p in [str(JOBS_ROOT), str(MODEL_ROOT), str(BIN_ROOT), str(Path.home())]:
        if p:
            s = s.replace(p, "[redacted]")
    return s


@spaces.GPU(duration=300)
def run_pipeline(file1, file2, dilate_prev_radius_vox=1, min_lesion_vol_ml=0.01):
    """
    file1: previous (baseline) FLAIR (.nii/.nii.gz or DICOM .zip)
    file2: new (follow-up) FLAIR (.nii/.nii.gz or DICOM .zip)
    Returns: (status_html, outputs_zip_path_or_None, report_html_update)
    """
    if file1 is None or file2 is None:
        return (
            "<div>⚠️ Please upload both the previous and the new scan.</div>",
            None,
            gr.update(value="", visible=False),
        )

    ensure_dirs()
    ensure_synthstrip_available()
    ensure_flames_model()

    job_dir = (JOBS_ROOT / f"msjob_{uuid.uuid4().hex}").resolve()
    job_dir.mkdir(parents=True, exist_ok=True)
    incoming_dir = job_dir / "incoming"
    incoming_dir.mkdir(parents=True, exist_ok=True)

    try:
        # Stage uploads
        prev_up = Path(file1.name)
        new_up  = Path(file2.name)
        prev_staged = stage_upload(prev_up, incoming_dir, "prev_upload")
        new_staged  = stage_upload(new_up,  incoming_dir, "new_upload")

        prev_is_zip = is_zip(prev_staged)
        new_is_zip  = is_zip(new_staged)

        # Convert ZIPs -> NIfTI if needed
        prev_nifti = convert_dicom_zip_to_nifti(prev_staged, job_dir) if prev_is_zip else prev_staged
        new_nifti  = convert_dicom_zip_to_nifti(new_staged,  job_dir) if new_is_zip  else new_staged

        # SynthStrip (NiPreps)
        stripped_dir = job_dir / STRIPPED_DIR
        prev_stripped, _ = skull_strip_synthstrip(prev_nifti, stripped_dir, "prev")
        new_stripped,  _ = skull_strip_synthstrip(new_nifti,  stripped_dir, "new")

        # Registration (new -> prev)
        reg_dir = job_dir / REGISTERED_DIR
        registered_path, _, _ = register_rigid_affine(prev_stripped, new_stripped, reg_dir)
        registered_path = registered_path.resolve()

        # FLAMeS segmentation
        seg_dir = job_dir / SEG_DIR; seg_dir.mkdir(parents=True, exist_ok=True)
        prev_mask_flames = seg_dir / "prev_flames_mask.nii.gz"
        new_mask_flames  = seg_dir / "new_in_prev_space_flames_mask.nii.gz"
        run_flames_single(prev_stripped,  prev_mask_flames, device="cuda")
        run_flames_single(registered_path, new_mask_flames,  device="cuda")

        # Diff maps
        diff_paths = build_diff_maps(
            prev_mask_flames, new_mask_flames, job_dir,
            dilate_prev_radius_vox=int(dilate_prev_radius_vox),
            min_lesion_vol_ml=float(min_lesion_vol_ml)
        )

        # Lesion counts for textual report
        new_lesions_count = count_lesions_in_mask(
            diff_paths["new_only"],
            min_lesion_vol_ml=0.0,             # already filtered in build_diff_maps
            use_26_connectivity=False
        )
        total_lesions_new_scan = count_lesions_in_mask(
            new_mask_flames,
            min_lesion_vol_ml=float(min_lesion_vol_ml),
            use_26_connectivity=False
        )

        # Package deliverables
        outputs_zip = package_selected(
            job_dir,
            prev_stripped,
            registered_path,
            diff_paths,
            prev_mask_flames=prev_mask_flames,
            new_mask_flames=new_mask_flames
        )

        if not outputs_zip.exists():
            raise RuntimeError("Packaging failed: outputs.zip not found.")

        # Textual report
        report_html = f"""
        <div>
          <h3>Textual report</h3>
          <ul>
            <li><strong>New lesions:</strong> {new_lesions_count}</li>
            <li><strong>Total lesions in new scan:</strong> {total_lesions_new_scan}</li>
          </ul>
          <div>
            Counts use the current minimum lesion volume filter of {float(min_lesion_vol_ml):.3f} mL (6-connectivity).
          </div>
        </div>
        """.strip()

        status_html = "✅ Done. Download the results below to inspect them in your preferred viewer."
        return (
            status_html,
            str(outputs_zip),
            gr.update(value=report_html, visible=True),
        )

    except Exception as e:
        msg = _redact_paths(str(e))
        status_html = f"⚠️ Error: {msg}"
        return (
            status_html,
            None,
            gr.update(value="", visible=False),
        )


# =========================
# Gradio UI (refined, consistent theme)
# =========================
with gr.Blocks(
    title=APP_NAME,
    analytics_enabled=False,
    theme=gr.themes.Default(),
    css="""
    /* ----- Title styling ----- */
    #title { 
        text-align: center; 
        margin-top: 1.5rem; 
        margin-bottom: 2rem;
        font-size: clamp(2rem, 3vw, 2.8rem);
        font-weight: 700;
        letter-spacing: -0.4px;
        background: linear-gradient(90deg, #4cafef, #7affd6);
        -webkit-background-clip: text;
        -webkit-text-fill-color: transparent;
    }

    /* ----- Run pipeline button ----- */
    #run_btn {
        display: block;
        margin: 28px auto 16px auto;
        background: linear-gradient(90deg, #4cafef, #7affd6);
        color: #000 !important;
        font-weight: 600 !important;
        font-size: 18px !important;
        border: none !important;
        border-radius: 10px !important;
        padding: 12px 28px !important;
        box-shadow: 0 4px 12px rgba(0, 0, 0, 0.15);
        transition: all 0.25s ease;
        text-transform: none !important;
    }

    #run_btn:hover {
        background: linear-gradient(90deg, #3da0df, #64eec8);
        transform: translateY(-1px);
        box-shadow: 0 6px 14px rgba(0, 0, 0, 0.18);
        color: #000
    }

    /* ----- Status box ----- */
    #status_box {
        overflow: visible !important;
        height: auto !important;
        max-height: none !important;
        min-height: 60px;
        padding: 14px 18px;
        border-radius: 10px;
        border: 1px solid var(--border-color-primary);
        background: var(--block-background-fill);
        font-size: 16px;
        line-height: 1.45;
    }

    /* ----- Info & reference sections ----- */
    .info-section {
        font-size: 18px;
        line-height: 1.7;
        max-width: 1000px;
        margin: 0 auto 28px auto;
        padding: 24px 30px;
        border-radius: 14px;
        background-color: var(--block-background-fill);
        border: 1px solid var(--border-color-primary);
        box-shadow: 0 4px 14px rgba(0,0,0,0.06);
    }

    .info-section h3 {
        margin-top: 0;
        margin-bottom: 12px;
        font-size: 22px;
        font-weight: 600;
        color: #4cafef;
        letter-spacing: -0.3px;
    }

    .info-section p, .info-section li {
        color: var(--body-text-color);
    }

    .info-section ul {
        margin-top: 6px;
        margin-bottom: 6px;
        padding-left: 24px;
        list-style-type: disc;
    }

    .info-section code { 
        background: var(--background-secondary);
        padding: 2px 5px;
        border-radius: 5px;
        font-size: 90%;
    }

    .info-section a {
        color: #4cafef;
        text-decoration: none;
    }

    .info-section a:hover {
        text-decoration: underline;
    }

    /* ----- Textual report styling ----- */
    #report {
        font-size: 20px !important;
        line-height: 1.8 !important;
        color: var(--body_text_color, var(--body-text-color));
        background: var(--block-background-fill);
        border: 1px solid var(--border-color-primary);
        border-radius: 16px;
        padding: 28px 34px;
        max-width: 1000px;
        margin: 0 auto 34px auto;
        box-shadow: 0 6px 18px rgba(0,0,0,0.08);
    }

    #report h3 {
        margin-top: 0;
        margin-bottom: 16px;
        font-size: 21px;
        font-weight: 700;
        color: #4cafef;
        letter-spacing: -0.3px;
        text-align: left;
    }

    #report ul {
        margin: 10px 0 0 22px;
        padding: 0;
        list-style-type: disc;
    }

    #report li {
        margin-bottom: 10px;
        font-size: 20px;
        line-height: 1.8;
    }

    #report strong {
        color: #4cafef;
        font-weight: 600;
        font-size: 20px;
    }

    #report .footnote {
        font-size: 16px;
        color: #999;
        margin-top: 14px;
    }
    """
) as demo:
    # ----- Title -----
    gr.Markdown(f"# {APP_NAME}", elem_id="title")

    # ----- Intro Section 1: Overview -----
    gr.HTML(
        """
<div class="info-section">
  <h3>Overview</h3>
  <p>This tool detects changes in <strong>multiple sclerosis (MS) lesions</strong> between two brain MRI scans.</p>

  <p>Input sequence must be <strong>isotropic 3D FLAIR</strong> in 
  <code>.nii/.nii.gz</code> (NIfTI) or DICOM (<code>.zip</code>) format. <br>
  If DICOM is provided, images are automatically converted to NIfTI using 
  <em>dcm2niix</em>.</p>

  <p>Processing includes skull stripping with <em>NiPreps SynthStrip</em> package, 
  rigid/affine co-registration of the two scans with SimpleITK, 
  and lesion segmentation using <em>FLAMeS</em> deep learning model.</p>

  <p>Lesion difference masks between the two scans are then calculated and made available for download.</p>

  <p><strong>Note: This application is a <em>research preview</em>. 
  For clinical reporting, all results should be reviewed and validated by a qualified radiologist.</strong></p>
</div>
"""
    )

    # ----- Intro Section 2: How to Use -----
    gr.HTML(
        """
<div class="info-section">
  <h3>How to use</h3>
  <ul>
    <li>Upload <em>previous (baseline)</em> and <em>new (follow-up)</em> 
        isotropic 3D FLAIR scans (<code>.nii/.nii.gz</code> or DICOM <code>.zip</code>).</li>
    <li>Click <strong>Run pipeline</strong>. Processing time takes approximately 3 minutes on current hardware.</li>
    <li>After processing, download the ZIP file and open the NIfTI outputs in your preferred neuroimaging viewer
        (e.g. ITK-SNAP, FSLeyes, 3D Slicer) to inspect the lesions and overlays.</li>
  </ul>

  <p style="margin-top:16px;"><strong>Advanced options:</strong></p>
  <ul>
    <li><em>Dilate previous mask (voxels):</em> Expands the baseline lesion mask slightly 
    to avoid missing small changes when comparing scans.</li>
    <li><em>Min lesion volume (mL):</em> Ignores very tiny spots below this volume, 
    so that noise is not counted as lesions.</li>
  </ul>
</div>
"""
    )

    # ----- File Inputs -----
    with gr.Row():
        prev_in = gr.File(label="Previous (baseline) FLAIR (.nii/.nii.gz or DICOM .zip)")
        new_in  = gr.File(label="New (follow-up) FLAIR (.nii/.nii.gz or DICOM .zip)")

    with gr.Accordion("Advanced options", open=False):
        dil    = gr.Slider(0, 3, value=1, step=1, label="Dilate previous mask (voxels)")
        minvol = gr.Slider(0.0, 0.2, value=0.01, step=0.005, label="Min lesion volume (mL)")

    # ----- Styled Run Button -----
    run_btn = gr.Button("Run pipeline", elem_id="run_btn")

    status  = gr.HTML(label="Status", elem_id="status_box")
    out_zip = gr.File(label="Download outputs (ZIP)")
    report  = gr.HTML(visible=False, label="Textual report", elem_id="report")

    run_btn.click(
        fn=run_pipeline,
        inputs=[prev_in, new_in, dil, minvol],
        outputs=[status, out_zip, report]
    )

    # ----- References -----
    gr.HTML(
        """
<div class="info-section">
<h3>References</h3>
<ol>
<li>
Li X, Morgan PS, Ashburner J, Smith J, Rorden C (2016). 
<em>The first step for neuroimaging data analysis: DICOM to NIfTI conversion.</em> 
<strong>J Neurosci Methods</strong> 264:47–56.  
<a href="https://doi.org/10.1016/j.jneumeth.2016.03.001" target="_blank">📄 DOI: 10.1016/j.jneumeth.2016.03.001</a>
</li>

<li>
Hoopes A, Mora JS, Dalca AV, Fischl B*, Hoffmann M* (2022). 
<em>SynthStrip: Skull-Stripping for Any Brain Image.</em> 
<strong>NeuroImage</strong> 260:119474.  
<a href="https://doi.org/10.1016/j.neuroimage.2022.119474" target="_blank">📄 DOI: 10.1016/j.neuroimage.2022.119474</a>
</li>

<li>
Dereskewicz E, La Rosa F, Dos Santos Silva J, et al. (2025). 
<em>FLAMeS: A Robust Deep Learning Model for Automated Multiple Sclerosis Lesion Segmentation.</em> 
<strong>medRxiv</strong>.  
<a href="https://doi.org/10.1101/2025.05.19.25327707" target="_blank">📄 DOI: 10.1101/2025.05.19.25327707</a>
</li>
</ol>
</div>
"""
    )


if __name__ == "__main__":
    demo.launch(
        server_name="0.0.0.0",
        server_port=7860,
        allowed_paths=[str(JOBS_ROOT)],
        show_error=True
    )