""" ================================================================================ VESUVIUS V7 - CONSERVATIVE IMPROVEMENT Based on V3 (0.461) with small incremental improvements: - Same 4x TTA (proven to work) - Same overlap 0.52 (proven to work) - Slightly lower thresholds for better recall - Better topology post-processing V3 settings: T_low=0.45, T_high=0.85, z_radius=1, xy_radius=0 V7 settings: T_low=0.42, T_high=0.82, z_radius=2, xy_radius=1 Target: Beat 0.461 with minimal risk ================================================================================ """ import subprocess import os import gc # Install packages var = "/kaggle/input/vesuvius25-packages-offline-installer-v20251226/whls" print(f"Installing from: {var}") subprocess.run([ "pip", "install", "--quiet", f"{var}/keras_nightly-3.12.0.dev2025100703-py3-none-any.whl", f"{var}/tifffile-2025.10.16-py3-none-any.whl", f"{var}/imagecodecs-2025.11.11-cp311-abi3-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", f"{var}/medicai-0.0.3-py3-none-any.whl", "--no-index", "--find-links", var ], check=False, capture_output=True) # Protobuf patch try: from google.protobuf import message_factory as _message_factory if not hasattr(_message_factory.MessageFactory, "GetPrototype"): from google.protobuf.message_factory import GetMessageClass def _GetPrototype(self, descriptor): return GetMessageClass(descriptor) _message_factory.MessageFactory.GetPrototype = _GetPrototype except: pass os.environ["KERAS_BACKEND"] = "jax" import keras from medicai.transforms import Compose, ScaleIntensityRange from medicai.models import TransUNet from medicai.utils.inference import SlidingWindowInference import numpy as np import pandas as pd import zipfile import tifffile import scipy.ndimage as ndi from skimage.morphology import remove_small_objects print(f"Backend: {keras.config.backend()}, Keras: {keras.version()}") # ============================================================================ # CONFIG - V7 CONSERVATIVE (Based on V3) # ============================================================================ root_dir = "/kaggle/input/vesuvius-challenge-surface-detection" test_dir = f"{root_dir}/test_images" output_dir = "/kaggle/working/submission_masks" zip_path = "/kaggle/working/submission.zip" os.makedirs(output_dir, exist_ok=True) # Model weights MODEL_PATH = "/kaggle/input/colab-a-162v4-gpu-transunet-seresnext101-x160/model.weights.h5" # V7 SETTINGS - Conservative improvements over V3 OVERLAP = 0.52 # Same as V3 (proven) # V7 threshold tuning (slightly more permissive than V3) # V3: T_low=0.45, T_high=0.85 # V7: T_low=0.42, T_high=0.82 (small decrease) T_LOW = 0.42 T_HIGH = 0.82 # V7 closing (slightly more aggressive for connectivity) # V3: z_radius=1, xy_radius=0 # V7: z_radius=2, xy_radius=1 Z_RADIUS = 2 XY_RADIUS = 1 # Dust removal (same as V3) DUST_MIN_SIZE = 100 # Load test data test_df = pd.read_csv(f"{root_dir}/test.csv") print(f"Test samples: {len(test_df)}") print(f"\n{'='*60}") print("V7 CONSERVATIVE - Incremental Improvement over V3") print(f"{'='*60}") print(f" - Overlap: {OVERLAP} (same as V3)") print(f" - TTA: 4x (same as V3)") print(f" - Threshold: T_low={T_LOW}, T_high={T_HIGH} (V3: 0.45, 0.85)") print(f" - Closing: z={Z_RADIUS}, xy={XY_RADIUS} (V3: 1, 0)") # ============================================================================ # TRANSFORMATION # ============================================================================ def val_transformation(image): data = {"image": image} pipeline = Compose([ ScaleIntensityRange( keys=["image"], a_min=0, a_max=255, b_min=0, b_max=1, clip=True, ), ]) return pipeline(data)["image"] # ============================================================================ # MODEL # ============================================================================ print(f"\nLoading model from: {MODEL_PATH}") model = TransUNet( input_shape=(160, 160, 160, 1), encoder_name='seresnext101', classifier_activation='softmax', num_classes=3, ) model.load_weights(MODEL_PATH) print(f"Model loaded! Params: {model.count_params() / 1e6:.1f}M") # Sliding window inference (same overlap as V3) pred_fn = SlidingWindowInference( model, roi_size=(160, 160, 160), num_classes=3, mode="gaussian", overlap=OVERLAP, sw_batch_size=1 ) # ============================================================================ # HELPERS (same as V3) # ============================================================================ def load_volume(path): vol = tifffile.imread(path) vol = vol.astype(np.float32) vol = vol[None, ..., None] return vol def rot90_volume(vol, k): if vol.ndim == 5: return np.rot90(vol, k=-k, axes=(2, 3)) else: return np.rot90(vol, k=-k, axes=(1, 2)) def unrot90_volume(vol, k): return rot90_volume(vol, (4 - k) % 4) def predict_probs_tta_rot(sample): """4x rotation TTA (same as V3)""" probs_accum = [] for k in range(4): s_rot = rot90_volume(sample, k) out = pred_fn(s_rot) out = np.asarray(out) probs = out[0, ..., 1] probs = unrot90_volume(probs, k) probs_accum.append(probs) return np.mean(probs_accum, axis=0) def build_anisotropic_struct(z_radius, xy_radius): z, r = z_radius, xy_radius if z == 0 and r == 0: return None depth = 2 * z + 1 if z > 0 else 1 size = 2 * r + 1 if r > 0 else 1 struct = np.zeros((depth, size, size), dtype=bool) cz = z if z > 0 else 0 cy = r if r > 0 else 0 cx = r if r > 0 else 0 for dz in range(-z if z > 0 else 0, z + 1 if z > 0 else 1): for dy in range(-r if r > 0 else 0, r + 1 if r > 0 else 1): for dx in range(-r if r > 0 else 0, r + 1 if r > 0 else 1): if r == 0 or dy * dy + dx * dx <= r * r: struct[cz + dz, cy + dy, cx + dx] = True return struct def topo_postprocess_v7(probs): """ V7: Improved post-processing - Slightly lower thresholds for better recall - More aggressive closing for better connectivity """ strong = probs >= T_HIGH weak = probs >= T_LOW if not strong.any(): return np.zeros_like(probs, dtype=np.uint8) # Hysteresis thresholding struct_hyst = ndi.generate_binary_structure(3, 3) mask = ndi.binary_propagation(strong, mask=weak, structure=struct_hyst) if not mask.any(): return np.zeros_like(probs, dtype=np.uint8) # Morphological closing (V7: slightly more aggressive) if Z_RADIUS > 0 or XY_RADIUS > 0: struct_close = build_anisotropic_struct(Z_RADIUS, XY_RADIUS) if struct_close is not None: mask = ndi.binary_closing(mask, structure=struct_close) # Remove small objects if DUST_MIN_SIZE > 0: mask = remove_small_objects(mask.astype(bool), min_size=DUST_MIN_SIZE) return mask.astype(np.uint8) def predict(sample): """V7 prediction pipeline""" probs_fg = predict_probs_tta_rot(sample) final = topo_postprocess_v7(probs_fg) return final # ============================================================================ # INFERENCE # ============================================================================ print(f"\n{'='*60}") print(f"V7 INFERENCE - {len(test_df)} images") print(f"{'='*60}") with zipfile.ZipFile(zip_path, "w", compression=zipfile.ZIP_DEFLATED) as z: for idx, image_id in enumerate(test_df["id"]): tif_path = f"{test_dir}/{image_id}.tif" print(f"\n[{idx+1}/{len(test_df)}] {image_id}") volume = load_volume(tif_path) volume = val_transformation(volume) output = predict(volume) out_path = f"{output_dir}/{image_id}.tif" tifffile.imwrite(out_path, output.astype(np.uint8)) z.write(out_path, arcname=f"{image_id}.tif") os.remove(out_path) print(f" Shape: {output.shape}, Voxels: {output.sum():,}") # Cleanup del volume, output gc.collect() print(f"\n{'='*60}") print(f"V7 CONSERVATIVE DONE! Submission: {zip_path}") print(f"{'='*60}")