"""Local DICOM viewing and inference adapter for the retained OCT submission. The Gradio controls and display behavior match OCT_submission. Prediction uses this project's saved model, normalization and postprocessing in an isolated worker; model weights and research records are unchanged. """ from __future__ import annotations import argparse import copy import html import json import math import os import shutil import signal import subprocess import sys import tempfile import time import uuid from collections import OrderedDict from dataclasses import dataclass, field from pathlib import Path from threading import Lock, RLock import numpy as np import pydicom import torch from PIL import Image, ImageDraw, ImageOps from src.io import source_frames, structural_header ROOT = Path(__file__).resolve().parents[1] CHECKPOINT = ROOT / "artifacts/checkpoint.pt" STRUCTURAL_OCT = "1.2.840.10008.5.1.4.1.1.77.1.5.4" FAMILIES = { "heidelberg_spectralis": "Heidelberg Spectralis", "topcon_maestro2": "Topcon Maestro2", "topcon_triton": "Topcon Triton", "zeiss_cirrus": "Zeiss Cirrus", } AUTO = "auto" NAMES = { "sct_280677004": "ILM", "dcm_128289": "RNFL · outer", "dcm_128290": "GCL · outer", "dcm_128291": "IPL · outer", "dcm_128292": "INL · outer", "dcm_128293": "OPL · outer", "sct_76710003": "ELM", "dcm_128320": "EZ · inner", "dcm_128321": "EZ · midline", "dcm_128323": "Interdigitation zone · inner", "dcm_128297": "RPE · inner", "dcm_128299": "RPE · outer", "dcm_128300": "Bruch’s membrane · outer", "dcm_128298": "RPE · centre", "dcm_128301": "Choroid–sclera interface", } PALETTE = [ "#22d3ee", "#a3e635", "#fbbf24", "#fb7185", "#c084fc", "#34d399", "#f472b6", "#60a5fa", "#f97316", "#e879f9", "#2dd4bf", "#fde047", "#a5b4fc", "#fda4af", "#d9f99d", ] COLORS = {key: PALETTE[i % len(PALETTE)] for i, key in enumerate(NAMES)} class ViewerError(ValueError): """A recoverable error suitable for displaying in the viewer.""" @dataclass class Prediction: positions: np.ndarray # K, number of predicted B-scans, original A keys: list[str] frames: list[int] family: str compute: str complete_volume: bool seconds: float export: Path @dataclass class Volume: identity: str directory: Path paths: list[Path] headers: list shape: tuple[int, int, int] detected_family: str | None axial_mm: float | None lateral_mm: float | None filename: str prediction: Prediction | None = None view_frame: int = 0 view_settings: tuple | None = None frames: OrderedDict = field(default_factory=OrderedDict) cancelled: bool = False updated: float = field(default_factory=time.monotonic) def frame(self, index: int) -> np.ndarray: index = max(0, min(int(index), self.shape[0] - 1)) self.updated = time.monotonic() if index not in self.frames: try: for path, header in self.headers: offset = int(header.get("ConcatenationFrameOffsetNumber", 0)) if offset <= index < offset + int(header.NumberOfFrames): pixels = np.asarray( next(source_frames(path, indices=[index - offset])) ) break else: raise IndexError("Frame outside volume") except Exception as exc: raise ViewerError( "Could not decode this DICOM. Install requirements.txt " "for JPEG2000 and lossless JPEG support. " + str(exc) ) from exc if pixels.shape != self.shape[1:]: raise ViewerError( "Decoded image dimensions do not match the DICOM header." ) self.frames[index] = pixels while len(self.frames) > 8: self.frames.popitem(last=False) self.frames.move_to_end(index) return self.frames[index] class SessionStore: def __init__(self): self.volumes: dict[str, Volume] = {} self.lock = RLock() def add(self, volume: Volume) -> str: with self.lock: self.reap() self.volumes[volume.identity] = volume return volume.identity def get(self, identity: str | None) -> Volume: with self.lock: if not identity or identity not in self.volumes: raise ViewerError("Upload a structural OCT DICOM first.") volume = self.volumes[identity] volume.updated = time.monotonic() return volume def remove(self, identity: str | None): with self.lock: volume = self.volumes.pop(identity, None) if volume: volume.cancelled = True shutil.rmtree(volume.directory, ignore_errors=True) def reap(self): for identity, volume in list(self.volumes.items()): if time.monotonic() - volume.updated > 7200: self.remove(identity) STORE = SessionStore() def detect_family(header) -> str | None: description = " ".join( str(header.get(tag, "")) for tag in ("Manufacturer", "ManufacturerModelName", "SeriesDescription") ).lower() if "spectralis" in description or "heidelberg" in description: return "heidelberg_spectralis" if "maestro" in description: return "topcon_maestro2" if "triton" in description: return "topcon_triton" if "cirrus" in description or "zeiss" in description: return "zeiss_cirrus" return None def pixel_spacing(header) -> tuple[float | None, float | None]: candidates = [] for key in ("SharedFunctionalGroupsSequence", "PerFrameFunctionalGroupsSequence"): sequence = header.get(key, []) if sequence: measures = sequence[0].get("PixelMeasuresSequence", []) if measures: candidates.append(measures[0].get("PixelSpacing")) candidates.append(header.get("PixelSpacing")) for values in candidates: if values is not None and len(values) == 2: row, col = map(float, values) if math.isfinite(row) and math.isfinite(col) and row > 0 and col > 0: return row, col return None, None def load_upload(files: list[str] | str) -> Volume: if isinstance(files, str): files = [files] if not files: raise ViewerError("Choose a .dcm file to begin.") directory = Path(tempfile.mkdtemp(prefix="myxos-oct-viewer-")) os.chmod(directory, 0o700) try: data = directory / "data" data.mkdir(mode=0o700) paths = [] all_headers = [] for index, uploaded in enumerate(files): source = Path(uploaded) if not source.is_file(): raise ViewerError( "An uploaded DICOM is no longer available. Upload it again." ) # A private copy decouples inference from Gradio's upload-cache lifetime. path = data / f"part-{index:03d}.dcm" shutil.copyfile(source, path) os.chmod(path, 0o600) header = pydicom.dcmread(path, stop_before_pixels=True) if str(header.get("SOPClassUID", "")) != STRUCTURAL_OCT: raise ViewerError( "This is not a structural OCT DICOM. Upload the OCT volume, rather than an OCTA or segmentation export." ) if int(header.get("SamplesPerPixel", 1)) != 1: raise ViewerError( "The selected checkpoint expects grayscale structural OCT." ) all_headers.append(header) paths.append(path) for tag in ("StudyInstanceUID", "SeriesInstanceUID", "ConcatenationUID"): identifiers = {str(h.get(tag)) for h in all_headers if h.get(tag)} if len(identifiers) > 1: raise ViewerError( "Upload one OCT volume, or the contiguous DICOM parts of that same volume." ) headers = [(path, structural_header(path)) for path in paths] headers.sort( key=lambda item: int(item[1].get("ConcatenationFrameOffsetNumber", 0)) ) count = 0 height, width = int(headers[0][1].Rows), int(headers[0][1].Columns) for path, header in headers: if int(header.get("ConcatenationFrameOffsetNumber", 0)) != count: raise ViewerError("Missing or overlapping concatenation frames.") if (int(header.Rows), int(header.Columns)) != (height, width): raise ViewerError("Inconsistent structural geometry.") if header.PhotometricInterpretation not in {"MONOCHROME1", "MONOCHROME2"}: raise ViewerError( "The selected checkpoint expects grayscale structural OCT." ) count += int(header.NumberOfFrames) if any( int(h.get("InConcatenationTotalNumber", len(paths))) != len(paths) for h in all_headers ): raise ViewerError("Upload all contiguous DICOM parts of this volume.") paths = [path for path, _ in headers] shape = count, height, width if not all(int(n) > 0 for n in shape): raise ViewerError("The DICOM contains an empty volume.") axial, lateral = pixel_spacing(all_headers[0]) volume = Volume( uuid.uuid4().hex, directory, paths, headers, tuple(shape), detect_family(all_headers[0]), axial, lateral, Path(files[0]).name, ) # Fail early when a required codec is missing. volume.frame((shape[0] - 1) // 2) return volume except ViewerError: shutil.rmtree(directory, ignore_errors=True) raise except Exception as exc: shutil.rmtree(directory, ignore_errors=True) raise ViewerError("Unable to load this OCT volume: " + str(exc)) from exc def resolve_family(volume: Volume, selected: str) -> str: family = volume.detected_family if selected == AUTO else selected if family not in FAMILIES: raise ViewerError( "The scanner could not be identified. Select its scanner family below." ) return family def boundary_choices(family: str | None) -> list[tuple[str, str]]: config = ENGINE.checkpoint.with_name("config.json") cfg = ( json.loads(config.read_text()) if config.is_file() else torch.load(ENGINE.checkpoint, map_location="cpu", weights_only=False)[ "config" ] ) keys = cfg["boundary_keys"].get(family, []) return [(NAMES.get(key, key), key) for key in keys] def compute_choices() -> list[tuple[str, str]]: choices = [("Automatic", "auto"), ("CPU", "cpu")] if ENGINE.checkpoint.suffix == ".npz": return choices if torch.cuda.is_available(): choices.append(("NVIDIA GPU · CUDA", "cuda")) if torch.backends.mps.is_available(): choices.append(("Apple GPU · MPS", "mps")) return choices def resolve_compute(selected: str) -> str: if selected == "auto": if ENGINE.checkpoint.suffix == ".npz": return "cpu" return ( "cuda" if torch.cuda.is_available() else ("mps" if torch.backends.mps.is_available() else "cpu") ) if selected not in {value for _, value in compute_choices()}: raise ViewerError("That compute device is not available on this machine.") return selected class InferenceEngine: def __init__(self): self.lock = Lock() self.checkpoint = CHECKPOINT self.threads = 2 self.decode_workers = 2 self.timeout = 1800 def predict(self, volume, family, compute, batch_size, full_volume, frame_index): device = resolve_compute(compute) frames = list(range(volume.shape[0])) if full_volume else [int(frame_index)] started = time.monotonic() volume.cancelled = False with ( self.lock, tempfile.TemporaryDirectory( prefix="inference-", dir=volume.directory ) as directory, ): work = Path(directory) job = { "checkpoint": str(self.checkpoint), "data_root": str(volume.directory / "data"), "oct_paths": [path.name for path in volume.paths], "family": family, "device": device, "batch_size": batch_size, "full_volume": full_volume, "frame_index": int(frame_index), "threads": self.threads, "decode_workers": self.decode_workers, "output": str(work / "prediction.npz"), "stop": str(work / "stop"), } spec = work / "job.json" spec.write_text(json.dumps(job)) log_path = work / "worker.log" with log_path.open("w") as log, log_path.open() as updates: with subprocess.Popen( [sys.executable, "-m", "src.viewer", "--job", str(spec)], cwd=ROOT, stdout=log, stderr=subprocess.STDOUT, start_new_session=os.name == "posix", ) as process: try: while True: if volume.cancelled: Path(job["stop"]).touch() for line in updates: try: event = json.loads(line) except json.JSONDecodeError: continue if isinstance(event, dict) and "completed" in event: volume.updated = time.monotonic() yield event try: returncode = process.wait(timeout=0.1) break except subprocess.TimeoutExpired: if time.monotonic() - started > self.timeout: raise ViewerError( f"Inference exceeded the {self.timeout}-second time limit." ) if volume.cancelled: raise ViewerError("Segmentation cancelled.") if returncode: detail = log_path.read_text()[-6000:] if "out of memory" in detail.lower(): raise ViewerError("GPU out of memory.") raise ViewerError( "Inference failed: " + detail.strip().splitlines()[-1] ) with np.load(job["output"], allow_pickle=False) as arrays: positions = arrays["positions"].copy() keys = arrays["boundary_keys"].tolist() if ( positions.shape != (len(keys), len(frames), volume.shape[2]) or not keys or not np.isfinite(positions).all() or positions.min() < 0 or positions.max() > volume.shape[1] - 1 ): raise ViewerError( "Inference returned invalid prediction geometry." ) suffix = ( "volume" if full_volume else f"preview-bscan-{frames[0] + 1:04d}" ) output = volume.directory / f"boundaries-{suffix}.npz" shutil.copyfile(job["output"], output) prediction = Prediction( positions, keys, frames, family, device, full_volume, time.monotonic() - started, output, ) volume.prediction = prediction yield {"result": prediction} finally: if process.poll() is None: if os.name == "posix": os.killpg(process.pid, signal.SIGKILL) else: process.kill() process.wait() ENGINE = InferenceEngine() def worker(job): """Run the saved pipeline with batch notifications, isolated from Gradio.""" from src.inference import apply_policy from src.inference.density import predict_bprob from src.inference.predict import ( ordered_process_prefetch, predict_prepared, prepare_pixels, ) from src.io import enforce_input_opens, resize_frames from src.models import make_model torch.set_num_threads(job["threads"]) row = {"case_id": "volume", "device": job["family"], "oct_paths": job["oct_paths"]} root = Path(job["data_root"]) started = time.monotonic() if Path(job["checkpoint"]).suffix == ".npz": from src.profile import load_checkpoint, predict_case models, cfg, _ = load_checkpoint(job["checkpoint"]) enforce_input_opens(root, [row]) positions, keys = predict_case(row, root, models, cfg, job["batch_size"]) if not job["full_volume"]: positions = positions[:, job["frame_index"] : job["frame_index"] + 1] else: saved = torch.load(job["checkpoint"], map_location="cpu", weights_only=False) if saved["config"].get("inference_wrapper") == "device_router_v1": route = saved["config"]["device_routes"][job["family"]] saved = saved["route_checkpoints"][route] cfg = saved["config"] device = torch.device(job["device"]) net = make_model(cfg) net.load_state_dict(saved["model"]) net.to(device).eval() audit = enforce_input_opens(root, [row]) selected = None if job["full_volume"]: prepared = ( list(ordered_process_prefetch([row], root, cfg, 1, audit))[0][1] if job["decode_workers"] else prepare_pixels(row, root, cfg) ) else: if cfg["architecture"] not in { "boundary_ensemble_v1", "boundary_unet_v1", "boundary_unet_axial_v1", "boundary_unet_spatial_v1", }: raise ViewerError("Use Full volume for this checkpoint architecture.") configs = [ cfg, *[component["config"] for component in cfg.get("components", [])], ] radius = max(c.get("context_radius", 0) for c in configs) headers = [ (root / name, structural_header(root / name)) for name in row["oct_paths"] ] count = sum(int(header.NumberOfFrames) for _, header in headers) index = job["frame_index"] lo, hi = max(0, index - radius), min(count, index + radius + 1) selected = index - lo frames = [] for path, header in headers: offset = int(header.get("ConcatenationFrameOffsetNumber", 0)) indices = range( max(0, lo - offset), min(int(header.NumberOfFrames), hi - offset) ) if indices: frames.extend(source_frames(path, indices=list(indices))) pixels = resize_frames(frames, cfg["height"], cfg["width"], "cpu").to( torch.float16 ) prepared = [(pixels, int(headers[0][1].Rows), int(headers[0][1].Columns))] cfg = copy.deepcopy(cfg) for c in [ cfg, *[component["config"] for component in cfg.get("components", [])], ]: if "density_b_sigma" in c: c["density_b_sigma"] = 0 policy = c.get("prediction_policy", {}) for settings in [policy, *policy.get("by_device", {}).values()]: settings.pop("median_b", None) settings.pop("gaussian_b_sigma", None) def report(completed, total): if Path(job["stop"]).exists(): raise ViewerError("Segmentation cancelled.") if selected is not None: completed, total = int(completed > selected), 1 print( json.dumps( { "completed": completed, "total": total, "device": job["device"], "seconds": time.monotonic() - started, } ), flush=True, ) if Path(job["stop"]).exists(): raise ViewerError("Segmentation cancelled.") if cfg.get("inference_wrapper") == "b_probability_smooth_v1": positions, keys = predict_bprob( row, prepared, net, cfg, device, job["batch_size"], progress=report ) else: positions, keys = predict_prepared( row, prepared, net, cfg, device, job["batch_size"], cfg.get("tta_horizontal", False), progress=report, ) positions = apply_policy(positions, job["family"], cfg.get("prediction_policy")) if selected is not None: positions = positions[:, selected : selected + 1] if Path(job["stop"]).exists(): raise ViewerError("Segmentation cancelled.") height = int(structural_header(root / row["oct_paths"][0]).Rows) arrays = { "positions": np.clip(positions, 0, height - 1).astype("float32"), "boundary_keys": np.asarray(keys), } if not job["full_volume"]: arrays["frame_indices"] = np.asarray([job["frame_index"]], dtype=np.int32) np.savez_compressed(job["output"], **arrays) def display_dimensions(volume: Volume, physical: bool) -> tuple[int, int]: _, h, a = volume.shape ratio = h / a if physical and volume.axial_mm and volume.lateral_mm: ratio *= volume.axial_mm / volume.lateral_mm width = 1000 height = max(1, round(width * ratio)) scale = min(1.0, 1400 / max(width, height)) return max(1, round(width * scale)), max(1, round(height * scale)) def render( volume: Volume, index: int, selected_keys: list[str] | None = None, overlay: bool = True, opacity: float = 0.9, line_width: int = 2, labels: bool = False, contrast: float = 1.0, brightness: float = 0.0, physical: bool = True, ) -> tuple[np.ndarray, np.ndarray | None]: index = max(0, min(volume.shape[0] - 1, int(index))) image = np.asarray(volume.frame(index), dtype=np.float32) finite = image[np.isfinite(image)] lo, hi = np.quantile(finite, [0.01, 0.995]) if finite.size else (0, 1) gray = np.nan_to_num((image - lo) / max(float(hi - lo), 1e-6)) gray = np.clip((gray - 0.5) * float(contrast) + 0.5 + float(brightness), 0, 1) width, height = display_dimensions(volume, physical) base = ( Image.fromarray(np.uint8(np.round(gray * 255))) .convert("RGB") .resize((width, height), Image.Resampling.BILINEAR) ) if volume.headers[0][1].PhotometricInterpretation == "MONOCHROME1": base = ImageOps.invert(base) raw = np.asarray(base) prediction = volume.prediction if prediction is None or index not in prediction.frames: return raw, None if not overlay: return raw, raw.copy() frame = prediction.frames.index(index) selected = set(prediction.keys if selected_keys is None else selected_keys) layer = base.convert("RGBA") annotation = Image.new("RGBA", base.size) draw = ImageDraw.Draw(annotation) _, native_h, native_a = volume.shape x = np.linspace(0, width - 1, native_a) for key, values in zip(prediction.keys, prediction.positions[:, frame]): if key not in selected: continue y = np.clip(values, 0, native_h - 1) * (height - 1) / max(1, native_h - 1) color = COLORS.get(key, "#ffffff") rgb = tuple(int(color[i : i + 2], 16) for i in (1, 3, 5)) points = list(zip(x.tolist(), y.tolist())) if len(points) > 1: draw.line( points, fill=(*rgb, round(255 * float(opacity))), width=max(1, int(line_width)), joint="curve", ) if labels: text = NAMES.get(key, key) anchor = max( 0, min( height - 14, round(float(np.median(y[: max(1, native_a // 8)])) - 13), ), ) box = draw.textbbox((8, anchor), text) draw.rectangle( (box[0] - 3, box[1] - 2, box[2] + 3, box[3] + 2), fill=(8, 15, 26, 210) ) draw.text((8, anchor), text, fill=(*rgb, 255)) return raw, np.asarray(Image.alpha_composite(layer, annotation).convert("RGB")) def legend(keys: list[str]) -> str: if not keys: return "Boundary colours appear after a scan is loaded." items = [] for key in keys: items.append( f'● {html.escape(NAMES.get(key, key))}' ) return " · ".join(items) def volume_info(volume: Volume, family: str | None) -> str: b, h, a = volume.shape scanner = FAMILIES.get(family, "Select scanner family") spacing = f" · axial {volume.axial_mm * 1000:.2f} µm" if volume.axial_mm else "" return ( f"**{html.escape(scanner)}** \n{b:,} B-scans · {h:,} × {a:,} pixels{spacing}" ) if __name__ == "__main__": parser = argparse.ArgumentParser(description="Isolated OCT viewer inference worker") parser.add_argument("--job", type=Path, required=True) worker(json.loads(parser.parse_args().job.read_text()))