fp2 / src /fiberpose /analysis.py
vikba's picture
Upload folder using huggingface_hub
beb4f15 verified
Raw History Blame Contribute Delete
28 kB
"""2D fiber analysis: DT prediction, centerlines, segmentation, and measurement sampling.
Every function is stateless: arrays in, results out. Per-image state lives on `Frame`.
"""
from __future__ import annotations
import logging
from typing import Any
import cv2
import numpy as np
from numba import njit
from scipy.ndimage import gaussian_filter, laplace, value_indices
from scipy.sparse import coo_matrix
from scipy.sparse.csgraph import connected_components
from skimage.segmentation import watershed
from .domain import FiberFragment
from .ml import OnnxFiberposeModel
from .raster import disk_mask, disk_radii, stamp_disks
from .ridge_detector import RidgeDetector
from .utils import (
_bilinear,
_sample_vec2,
normals_from_dt_2d,
resample_polyline_xy,
)
logger = logging.getLogger(__name__)
Sample = tuple[int, int, float, float, float, float]
# ------------------------------------------------------------------ #
# Public pipeline stages
def predict(model: OnnxFiberposeModel, image: np.ndarray) -> dict[str, np.ndarray | None]:
"""Predict the DT of `image` plus its derived fields, keyed like `Frame.set_prediction_outputs`."""
h, w = image.shape[:2]
stride = int(model.stride)
pad_width = ((0, -h % stride), (0, -w % stride)) + ((0, 0),) * (image.ndim - 2)
dt_pad, fg_pad = model.predict(np.pad(image, pad_width, mode="reflect"))
dt = np.asarray(dt_pad[:h, :w], dtype=np.float32)
dt_smooth = gaussian_filter(dt, sigma=1.5)
gy, gx = np.gradient(dt_smooth)
return {
"dt": dt,
"dt_u8": np.clip(dt, 0, 255).astype(np.uint8),
"dt_normalized": dt / 255.0,
"foreground_prob": None if fg_pad is None else np.asarray(fg_pad[:h, :w], dtype=np.float32),
"gx": np.asarray(gaussian_filter(gx, sigma=1.0), dtype=np.float32),
"gy": np.asarray(gaussian_filter(gy, sigma=1.0), dtype=np.float32),
"lapl_img": np.asarray(laplace(dt_smooth), dtype=np.float32),
}
def segment(
dt: np.ndarray,
dt_u8: np.ndarray,
gx: np.ndarray,
gy: np.ndarray,
foreground_prob: np.ndarray | None,
*,
segmentation_mode: str,
watershed_grad_threshold: float,
min_segment_foreground_mean: float,
max_angle_deg: float,
ridge_sigma: float,
ridge_low: float,
ridge_high: float,
strict_centerlines: bool,
keep_segments_without_measurement_centerlines: bool,
merge_expansion_factor: float,
segmentation_expansion_factor: float,
min_fragment_centerline_length_px: float,
measurement_centerline_length_trim_fraction: float,
grad_quality_trim_fraction: float,
) -> tuple[np.ndarray, list[FiberFragment]]:
"""Segment fibers: ridges -> fragments -> uint16 labels.
Returns the labels and the fragments that still own a label after pruning.
"""
shape = dt_u8.shape
detected_polylines = RidgeDetector.detect_ridges(
image2d=dt_u8,
sigma=float(ridge_sigma),
low=float(ridge_low),
high=float(ridge_high),
polarity="light",
)
raw_fragments = _group_detected_polylines_into_segmentation_fragments(
detected_polylines=detected_polylines,
dt=dt,
merge_expansion_factor=float(merge_expansion_factor),
min_fragment_centerline_length_px=float(min_fragment_centerline_length_px),
trim=3,
)
fragments = _filter_measurement_centerlines(
segmentation_fragments=raw_fragments,
dt_u8=dt_u8,
dt=dt,
gx=gx,
gy=gy,
max_angle_deg=max_angle_deg,
win_len=25.0,
std_thresh=0.8,
strict=strict_centerlines,
keep_segments_without_measurement_centerlines=keep_segments_without_measurement_centerlines,
)
# Drop the fragments with the shortest measurement centerlines.
trim_count = int(np.floor(len(fragments) * _clamp_trim_fraction(measurement_centerline_length_trim_fraction)))
if trim_count > 0:
ordered = sorted(
fragments,
key=lambda seg: (float(seg.total_measurement_centerline_length_px), int(seg.fragment_id)),
)
drop_ids = {int(seg.fragment_id) for seg in ordered[:trim_count]}
fragments = [seg for seg in fragments if int(seg.fragment_id) not in drop_ids]
grad_err = np.abs(np.sqrt(gx * gx + gy * gy) - 1.0)
# Pixels each fragment covers at the segmentation expansion, shared by both steps below.
fragment_pixels = _fragment_pixels(fragments, float(segmentation_expansion_factor), shape)
fragments, fragment_pixels = _filter_fragments_by_grad_quality(
fragments=fragments,
fragment_pixels=fragment_pixels,
grad_err=grad_err,
trim_fraction=_clamp_trim_fraction(grad_quality_trim_fraction),
min_valid_pixels=8,
)
if segmentation_mode == "radial_expansion":
labels_i32 = _rasterize_segmentation_fragments(fragments=fragments, fragment_pixels=fragment_pixels, shape=shape)
labels = np.clip(labels_i32, 0, np.iinfo(np.uint16).max).astype(np.uint16, copy=False)
else:
markers = _render_markers(shape, fragments)
labels = _watershed(dt_u8, gx, gy, markers, gradient_threshold=watershed_grad_threshold)
labels = _prune_labels_by_foreground(
labels,
foreground_prob=foreground_prob,
min_segment_foreground_mean=min_segment_foreground_mean,
)
surviving_ids = set(np.unique(labels)) - {0}
fragments = [fragment for fragment in fragments if int(fragment.parent_fiber_id) in surviving_ids]
logger.info("Segmented %d fragments (%d detected)", len(fragments), len(raw_fragments))
return labels, fragments
def sample_centerlines(
dt: np.ndarray,
gx: np.ndarray,
gy: np.ndarray,
fragments: list[FiberFragment],
spacing_px: float = 5.0,
scale_factor: float = 1.0,
) -> tuple[list[Sample], list[Sample]]:
"""Sample thickness and orientation along the fragments' measurement centerlines.
Returns:
- measurement_samples: (y, x, ny, nx, r, angle_deg) in original/display scale
- drawing_samples: (y, x, ny, nx, r, angle_deg) in processing-image scale
Note:
`_xy` suffix denotes (x, y) ordering, while `_yx` denotes (y, x).
"""
spacing = max(1.0, float(spacing_px))
scale = float(scale_factor)
if not np.isfinite(scale) or scale <= 0:
scale = 1.0
measurement_samples: list[Sample] = []
drawing_samples: list[Sample] = []
def _canon_angle(ty: float, tx: float) -> float:
n = float(np.hypot(ty, tx))
if n == 0.0 or not np.isfinite(n):
return float("nan")
ty, tx = ty / n, tx / n
ang = -np.degrees(np.arctan2(ty, tx))
if ang > 90.0:
ang -= 180.0
elif ang <= -90.0:
ang += 180.0
if ang <= -90.0:
ang = -89.999
return float(ang)
for frag in fragments:
for seg_xy in frag.measurement_centerlines_xy:
samples_xy = resample_polyline_xy(seg_xy.astype(np.float32), step=spacing)
if len(samples_xy) < 2:
continue
samples_yx = samples_xy[:, [1, 0]].astype(np.float32)
normals_yx = normals_from_dt_2d(
centerline=samples_yx,
dt=dt,
gx=gx,
gy=gy,
alpha=0.5,
iters=3,
tau_g=1e-3,
)
for i, (y, x) in enumerate(samples_yx):
ny, nx = float(normals_yx[i, 0]), float(normals_yx[i, 1])
norm = float(np.hypot(ny, nx))
if norm == 0.0 or not np.isfinite(norm):
continue
ny /= norm
nx /= norm
angle_deg = _canon_angle(nx, -ny)
v1 = float(_bilinear(dt, y - ny, x - nx))
v2 = float(_bilinear(dt, y + ny, x + nx))
r = 0.5 * (v1 + v2) + 1.0
if r <= 0 or not np.isfinite(r):
continue
drawing_samples.append((int(round(y)), int(round(x)), ny, nx, r, angle_deg))
measurement_samples.append(
(int(round(y / scale)), int(round(x / scale)), ny, nx, r / scale, angle_deg)
)
return measurement_samples, drawing_samples
def void_mask_and_pore_dt(dt_u8: np.ndarray, thr_u8: int) -> tuple[np.ndarray, np.ndarray]:
"""
Returns:
void_mask_u8: uint8, {0,1}, where 1 marks void (pores), opened to drop thin artifacts
pore_dt: float32, Euclidean distance (in pixels) inside voids
"""
t = float(thr_u8) / 255.0
dt01 = dt_u8.astype(np.float32) / 255.0
void_mask = (~(dt01 >= t)).astype(np.uint8)
if void_mask.any():
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
void_mask = cv2.morphologyEx(void_mask, cv2.MORPH_OPEN, kernel)
if void_mask.any():
pore_dt = cv2.distanceTransform(void_mask, distanceType=cv2.DIST_L2, maskSize=5).astype(np.float32)
else:
pore_dt = np.zeros_like(dt01, dtype=np.float32)
return void_mask, pore_dt
def sample_inscribed_circles(
pore_dt: np.ndarray,
void_mask: np.ndarray,
scale_factor: float = 1.0,
min_radius_px: float = 1.0,
min_peak_px: float = 3.0,
peak_footprint: int = 3,
) -> list[dict[str, Any]]:
"""Greedily place non-overlapping disks at pore-DT peaks, largest first."""
scale = scale_factor or 1.0
if not np.isfinite(scale) or scale <= 0:
scale = 1.0
H, W = pore_dt.shape[:2]
current_void = (void_mask.astype(np.uint8) > 0).astype(np.uint8)
# Local maxima inside the void. Zeros become 255 so background never counts as a peak.
k = max(3, int(peak_footprint) | 1)
arr_work = pore_dt.copy()
arr_work[arr_work == 0] = 255
dil = cv2.dilate(arr_work, cv2.getStructuringElement(cv2.MORPH_RECT, (k, k)))
ys, xs = np.nonzero((arr_work == dil) & (pore_dt > 0) & current_void.astype(bool))
if xs.size == 0:
return []
radii = pore_dt[ys, xs]
keep = radii >= float(min_peak_px)
if not np.any(keep):
return []
xs, ys, radii = xs[keep], ys[keep], radii[keep]
order = np.argsort(-radii)
xs, ys, radii = xs[order], ys[order], radii[order]
circles: list[dict[str, Any]] = []
for x, y, r in zip(xs, ys, radii):
if r < float(min_radius_px):
continue
r_int = int(np.floor(float(r)))
if r_int <= 0:
continue
if x - r_int < 0 or x + r_int >= W or y - r_int < 0 or y + r_int >= H:
continue
# The disk lies fully inside the image: test and carve only its window.
window = current_void[y - r_int:y + r_int + 1, x - r_int:x + r_int + 1]
disk = disk_mask(r_int)
if window[disk].all():
circles.append({'x': int(x), 'y': int(y), 'r_px': float(r), 'r_px_orig': float(r) / scale})
window[disk] = 0
return circles
# ------------------------------------------------------------------ #
# Fragments
def _group_detected_polylines_into_segmentation_fragments(
*,
detected_polylines: list[np.ndarray],
dt: np.ndarray,
merge_expansion_factor: float = 0.7,
min_fragment_centerline_length_px: float = 10.0,
trim: int = 3,
) -> list[FiberFragment]:
"""
Group detected polylines into segmentation fragments using overlap under merge expansion.
"""
h, w = dt.shape[:2]
trim_n = max(0, int(trim))
polylines: list[np.ndarray] = []
polyline_radii: list[np.ndarray] = []
for pts_xy_raw in detected_polylines:
pts_xy = np.asarray(pts_xy_raw, dtype=np.float32)
if pts_xy.ndim != 2 or pts_xy.shape[1] != 2:
continue
if trim_n > 0 and pts_xy.shape[0] > 2 * trim_n:
pts_xy = pts_xy[trim_n:-trim_n]
if pts_xy.shape[0] < 2:
continue
x_idx = np.clip(np.rint(pts_xy[:, 0]).astype(np.int32), 0, w - 1)
y_idx = np.clip(np.rint(pts_xy[:, 1]).astype(np.int32), 0, h - 1)
polylines.append(pts_xy)
polyline_radii.append(np.asarray(dt[y_idx, x_idx], dtype=np.float32))
if not polylines:
return []
group_ids = _overlap_groups(polylines, polyline_radii, max(0.0, float(merge_expansion_factor)), (h, w))
grouped_polylines: dict[int, list[np.ndarray]] = {}
grouped_radii: dict[int, list[np.ndarray]] = {}
for label_id, pts_xy, radii in zip(group_ids.tolist(), polylines, polyline_radii):
grouped_polylines.setdefault(label_id, []).append(pts_xy)
grouped_radii.setdefault(label_id, []).append(radii)
fragments: list[FiberFragment] = []
for label_id in sorted(grouped_polylines):
max_centerline_length = max(
FiberFragment._centerline_length(pts_xy) for pts_xy in grouped_polylines[label_id]
)
if max_centerline_length < float(min_fragment_centerline_length_px):
continue
fragments.append(
FiberFragment(
fragment_id=len(fragments) + 1,
segment_centerlines_xy=grouped_polylines[label_id],
segment_radii=grouped_radii[label_id],
parent_fiber_id=int(label_id),
)
)
return fragments
def _overlap_groups(
polylines: list[np.ndarray],
radii: list[np.ndarray],
expansion: float,
shape: tuple[int, int],
) -> np.ndarray:
"""Group ids (1-based) for polylines whose expanded disks share at least one pixel.
Groups are the connected components of the "shares a pixel" graph, numbered in
order of their first polyline.
"""
n = len(polylines)
owners = np.repeat(np.arange(n), [len(p) for p in polylines])
pixel, owner = stamp_disks(
np.concatenate(polylines), disk_radii(np.concatenate(radii), expansion), owners, shape
)
# Sort by pixel; consecutive entries on the same pixel link two polylines.
order = np.lexsort((owner, pixel))
pixel, owner = pixel[order], owner[order]
shared = pixel[1:] == pixel[:-1]
edges = np.unique(owner[:-1][shared] * n + owner[1:][shared])
graph = coo_matrix((np.ones(edges.size, dtype=np.int32), (edges // n, edges % n)), shape=(n, n))
_, component = connected_components(graph, directed=False)
_, first_polyline = np.unique(component, return_index=True)
rank = np.empty_like(first_polyline)
rank[np.argsort(first_polyline)] = np.arange(first_polyline.size)
return rank[component] + 1
def _fragment_pixels(
fragments: list[FiberFragment],
segmentation_expansion_factor: float,
shape: tuple[int, int],
) -> list[np.ndarray]:
"""Flat pixel indices each fragment covers: disks of its local radius along its centerlines."""
if not fragments:
return []
points, radii, owners = [], [], []
for index, frag in enumerate(fragments):
for pts_xy, frag_radii in zip(frag.segment_centerlines_xy, frag.segment_radii):
points.append(pts_xy)
radii.append(frag_radii)
owners.append(np.full(len(pts_xy), index))
expansion = max(0.0, float(segmentation_expansion_factor))
pixel, owner = stamp_disks(
np.concatenate(points), disk_radii(np.concatenate(radii), expansion), np.concatenate(owners), shape
)
# stamp_disks sorts by owner, so each fragment's pixels are one contiguous run.
return np.split(pixel, np.cumsum(np.bincount(owner, minlength=len(fragments)))[:-1])
def _rasterize_segmentation_fragments(
*,
fragments: list[FiberFragment],
fragment_pixels: list[np.ndarray],
shape: tuple[int, int],
) -> np.ndarray:
"""Render radial-expansion labels; where fragments overlap, the later one wins."""
expanded_labels = np.zeros(shape, dtype=np.int32)
flat = expanded_labels.reshape(-1)
for frag, pixels in zip(fragments, fragment_pixels):
flat[pixels] = int(frag.parent_fiber_id)
return expanded_labels
def _filter_fragments_by_grad_quality(
*,
fragments: list[FiberFragment],
fragment_pixels: list[np.ndarray],
grad_err: np.ndarray,
trim_fraction: float,
min_valid_pixels: int = 8,
) -> tuple[list[FiberFragment], list[np.ndarray]]:
"""Score fragments by median |grad(dt)| error off their centerlines; drop unreliable/worst ones.
Returns the kept fragments and their pixel lists, in input order.
"""
h, w = grad_err.shape[:2]
grad_err_flat = grad_err.reshape(-1)
reliable: list[tuple[FiberFragment, np.ndarray]] = []
for fragment, pixels in zip(fragments, fragment_pixels):
centerline_pixels = [
np.clip(np.rint(pts_xy[:, 1]).astype(np.int64), 0, h - 1) * w
+ np.clip(np.rint(pts_xy[:, 0]).astype(np.int64), 0, w - 1)
for pts_xy in fragment.segment_centerlines_xy
]
values = grad_err_flat[pixels[~np.isin(pixels, np.concatenate(centerline_pixels))]]
values = np.asarray(values[np.isfinite(values)], dtype=np.float32)
score = float(np.median(values)) if values.size else float("nan")
fragment.grad_quality_score = score
if values.size >= int(min_valid_pixels) and np.isfinite(score):
reliable.append((fragment, pixels))
drop_count = int(np.floor(len(reliable) * float(trim_fraction)))
if drop_count > 0:
ordered = sorted(reliable, key=lambda item: (-float(item[0].grad_quality_score), int(item[0].fragment_id)))
drop_ids = {int(frag.fragment_id) for frag, _ in ordered[:drop_count]}
reliable = [item for item in reliable if int(item[0].fragment_id) not in drop_ids]
return [frag for frag, _ in reliable], [pixels for _, pixels in reliable]
def _filter_measurement_centerlines(
*,
segmentation_fragments: list[FiberFragment],
dt_u8: np.ndarray,
dt: np.ndarray,
gx: np.ndarray | None,
gy: np.ndarray | None,
max_angle_deg: float,
win_len: float,
std_thresh: float,
strict: bool,
keep_segments_without_measurement_centerlines: bool,
) -> list[FiberFragment]:
"""Collect measurement centerlines per fragment and update fragments in place."""
if strict and (gx is None or gy is None):
raise ValueError("Strict centerline filtering requires dt, grad_x, and grad_y.")
if not strict:
gx = gy = dt # never read outside strict mode, but numba needs arrays
active_segmentation_fragments: list[FiberFragment] = []
for seg in segmentation_fragments:
seg.measurement_centerlines_xy = []
for polyline_xy in seg.segment_centerlines_xy:
pts_xy = np.ascontiguousarray(polyline_xy, dtype=np.float32).reshape(-1, 2)
spans = _window_spans(
pts_xy, dt_u8, dt, gx, gy, float(win_len), float(std_thresh), float(max_angle_deg), bool(strict)
)
seg.measurement_centerlines_xy += [pts_xy[s:e + 1].copy() for s, e in spans]
seg.recompute_total_measurement_centerline_length_px()
if seg.has_measurement_centerline:
active_segmentation_fragments.append(seg)
if bool(keep_segments_without_measurement_centerlines):
active_segmentation_fragments = list(segmentation_fragments)
return active_segmentation_fragments
def _clamp_trim_fraction(value: float) -> float:
value = float(value)
if not np.isfinite(value):
return 0.05
return float(min(0.95, max(0.0, value)))
# ------------------------------------------------------------------ #
# Labels
def _watershed(
dt_u8: np.ndarray,
gx: np.ndarray,
gy: np.ndarray,
markers: np.ndarray,
gradient_threshold: float,
) -> np.ndarray:
"""
Watershed on inverted DT, constrained by thresholded gradient magnitude.
Two dilations of the detected centerline markers are added to the
threshold mask. The combined mask is then eroded and dilated before
watershed.
"""
dt_u8 = np.asarray(dt_u8)
gx = np.asarray(gx, dtype=np.float32)
gy = np.asarray(gy, dtype=np.float32)
markers = np.asarray(markers)
if gx.shape != dt_u8.shape or gy.shape != dt_u8.shape or markers.shape != dt_u8.shape:
raise ValueError("DT, gradients, and markers must have matching shapes.")
if not markers.any():
return np.zeros_like(dt_u8, dtype=np.uint16)
threshold = float(gradient_threshold)
if not np.isfinite(threshold) or threshold < 0.0:
raise ValueError("Gradient magnitude threshold must be finite and non-negative.")
dt_norm = dt_u8.astype(np.float32) / 255.0
grad_magnitude = np.hypot(gx, gy)
fiber_region_mask = np.isfinite(grad_magnitude) & (grad_magnitude >= threshold)
morphology_kernel = np.ones((3, 3), dtype=np.uint8)
centerline_mask = (markers > 0).astype(np.uint8)
dilated_centerlines = cv2.dilate(
centerline_mask,
morphology_kernel,
iterations=2,
).astype(bool)
fiber_region_mask |= dilated_centerlines
fiber_region_mask_u8 = fiber_region_mask.astype(np.uint8)
fiber_region_mask_u8 = cv2.erode(
fiber_region_mask_u8,
morphology_kernel,
iterations=1,
)
fiber_region_mask = cv2.dilate(
fiber_region_mask_u8,
morphology_kernel,
iterations=1,
).astype(bool)
elev = 1.0 - dt_norm
if not fiber_region_mask.any():
return np.zeros_like(dt_u8, dtype=np.uint16)
marker_pixels = markers > 0
logger.info(
"Watershed gradient mask: threshold=%.4f mask=%d/%d marker_pixels=%d/%d",
threshold,
int(np.count_nonzero(fiber_region_mask)),
int(fiber_region_mask.size),
int(np.count_nonzero(fiber_region_mask & marker_pixels)),
int(np.count_nonzero(marker_pixels)),
)
conn8 = np.ones((3, 3), dtype=bool)
labels = watershed(
elev,
markers=markers,
mask=fiber_region_mask,
connectivity=conn8,
).astype(np.uint16)
return labels
def _prune_labels_by_foreground(
labels: np.ndarray,
*,
foreground_prob: np.ndarray | None,
min_segment_foreground_mean: float = 0.5,
) -> np.ndarray:
if foreground_prob is None:
return labels
labels_arr = np.asarray(labels)
foreground_arr = np.asarray(foreground_prob, dtype=np.float32)
if labels_arr.shape != foreground_arr.shape:
raise ValueError("Foreground probability must match label shape.")
threshold = float(min_segment_foreground_mean)
if not np.isfinite(threshold):
threshold = 0.5
threshold = float(np.clip(threshold, 0.0, 1.0))
pruned = labels_arr.copy()
# value_indices lists each label's pixels in raster order, so the float32 mean
# sums the same values in the same order as a boolean mask would.
for label_id, pixels in value_indices(labels_arr).items():
if label_id > 0 and float(np.mean(foreground_arr[pixels])) < threshold:
pruned[pixels] = 0
return pruned.astype(labels_arr.dtype, copy=False)
def _render_markers(shape: tuple[int, int], segments: list[FiberFragment]) -> np.ndarray:
markers = np.zeros(shape, dtype=np.int32)
for seg in segments:
for polyline_xy in seg.segment_centerlines_xy:
pts = np.asarray(polyline_xy, dtype=np.float32)
pts_int = np.round(pts).astype(np.int32).reshape(-1, 1, 2)
cv2.polylines(
markers,
[pts_int],
isClosed=False,
color=int(seg.parent_fiber_id),
thickness=1,
lineType=cv2.LINE_8,
)
return markers
# ------------------------------------------------------------------ #
# Measurement-centerline windows
# Strict-mode thresholds; never configured, so numba folds them in as constants.
_GRAD_ALIGNMENT_COS = 0.95
_RADIUS_REL_TOL = 0.25
_MIN_VALID_FRACTION = 0.7
@njit(cache=True)
def _window_spans(pts_xy, image, dt, gx, gy, win_len, std_thresh, max_angle_deg, strict):
"""Merged (start, end) point spans of flat, level windows, each at least `win_len` long.
Windows of arc length `win_len` slide along the polyline in steps of `win_len / 2`.
A window is accepted when `image` is flat along it (low std, small end-to-end rise)
and, in strict mode, when the DT looks like a clean fiber cross-section.
Overlapping accepted windows merge as they are found.
"""
n = len(pts_xy)
h, w = image.shape
cum = np.zeros(n, dtype=np.float32) # arc length at each point
intens = np.empty(n, dtype=np.float64) # image at each point, rounded to the nearest pixel
total = 0.0
for i in range(n):
if i > 0:
dx, dy = pts_xy[i, 0] - pts_xy[i - 1, 0], pts_xy[i, 1] - pts_xy[i - 1, 1]
total += np.sqrt(dx * dx + dy * dy)
cum[i] = total
x = min(max(int(np.rint(pts_xy[i, 0])), 0), w - 1)
y = min(max(int(np.rint(pts_xy[i, 1])), 0), h - 1)
intens[i] = image[y, x]
spans = np.empty((n, 2), dtype=np.int64)
k = 0
if n < 2 or cum[-1] <= 0:
return spans[:0]
s = 0
while s < n - 1:
e = min(max(np.searchsorted(cum, cum[s] + win_len), s + 1), n - 1)
if _is_flat(intens[s:e + 1], cum[e] - cum[s], std_thresh, max_angle_deg) and (
not strict or _is_clean_cross_section(pts_xy[s:e + 1], dt, gx, gy)
):
if k > 0 and s <= spans[k - 1, 1]:
spans[k - 1, 1] = max(spans[k - 1, 1], e)
else:
spans[k, 0], spans[k, 1] = s, e
k += 1
next_s = np.searchsorted(cum, cum[s] + win_len / 2.0)
if next_s <= s:
break
s = next_s
spans = spans[:k]
return spans[cum[spans[:, 1]] - cum[spans[:, 0]] >= win_len]
@njit(cache=True)
def _is_flat(intens, length, std_thresh, max_angle_deg):
"""Low intensity spread, and a rise from first to last point below `max_angle_deg`.
Intensities are integers, so distinct stds are far apart and float32 rounding only
merges a std with a threshold it exactly equals; that tie counts as not flat.
"""
if length <= 0.0 or np.float32(np.sqrt(np.mean((intens - np.mean(intens)) ** 2))) >= np.float32(std_thresh):
return False
return np.degrees(np.arctan(abs(intens[-1] - intens[0]) / length)) < max_angle_deg
@njit(cache=True)
def _is_clean_cross_section(pts_xy, dt, gx, gy):
"""Most points look like a clean fiber cross-section in the DT.
At each point, probe half the local radius r to either side along the normal: the
DT gradients there must point in opposite directions and the DT must be about r/2.
"""
n = len(pts_xy)
valid = total = 0
for i in range(n):
a, b = max(i - 1, 0), min(i + 1, n - 1)
tx, ty = pts_xy[b, 0] - pts_xy[a, 0], pts_xy[b, 1] - pts_xy[a, 1]
tan_norm = np.sqrt(tx * tx + ty * ty) + np.float32(1e-6)
nx, ny = float(-ty / tan_norm), float(tx / tan_norm)
px, py = float(pts_xy[i, 0]), float(pts_xy[i, 1])
r = _bilinear(dt, py, px)
if r <= 1e-6 or (abs(nx) <= 1e-8 and abs(ny) <= 1e-8):
continue
h = 0.5 * r
g_plus = _sample_vec2(gx, gy, py + ny * h, px + nx * h)
g_minus = _sample_vec2(gx, gy, py - ny * h, px - nx * h)
mag_plus = np.sqrt(g_plus[0] ** 2 + g_plus[1] ** 2)
mag_minus = np.sqrt(g_minus[0] ** 2 + g_minus[1] ** 2)
if mag_plus <= 1e-6 or mag_minus <= 1e-6:
continue
u_plus, u_minus = g_plus / (mag_plus + 1e-6), g_minus / (mag_minus + 1e-6)
opposed = -(u_plus[0] * u_minus[0] + u_plus[1] * u_minus[1]) >= _GRAD_ALIGNMENT_COS
tol = _RADIUS_REL_TOL * max(r, 1e-6)
radius_ok = (
abs(_bilinear(dt, py + ny * h, px + nx * h) - h) <= tol
and abs(_bilinear(dt, py - ny * h, px - nx * h) - h) <= tol
)
total += 1
valid += opposed and radius_ok
return total > 0 and valid / total >= _MIN_VALID_FRACTION