Skip to main content

functions

Pure per-scan compute for the lesion calculation step.

Version-agnostic: parses one scan's B-scan predictions once, measures every configured group (area, volume, per-lesion detail, fovea-relative geometry), and lifts the result into the cache-facing LesionMetrics dataclass.

Module​

Functions​

compute_lesion_metrics_for_scan​

def compute_lesion_metrics_for_scan(    bscan_predictions: tuple[str, ...],    groups: Mapping[str, MaskGroupSpec],    scale: VoxelScale,    *,    fovea_centre: tuple[int, int, int] | None = None,    roi_radius_mm: float | None = None,    max_lesions_per_group: int = 256,    declared_num_bscans: int | None = None,    angular_span_min_radius_mm: float | None = None,) ‑> LesionMetrics:

Measure every configured group for one scan.

Every area-eligible segmentation is parsed (not only the labels the configured groups name), because parse_mask_json requires a contiguous label-index space sized to the mapping it is given. The groups are handed to the parse rather than applied after it, because a group's voxel count has to be accumulated while each B-scan's raster still carries axial positions — combining per-label counts afterwards cannot recover them. Depth is collected only when scale supports a volume, so a group with no axial spacing pays nothing for a measurement it cannot report, and only for groups whose volume is meaningful (group_reports_volume).

Arguments

  • bscan_predictions: One raw prediction string per B-scan.
  • groups: Named label groups to measure.
  • scale: The cube's voxel dimensions.
  • fovea_centre: (bscan, column, row) voxel index of the fovea landmark, or None when no landmark was available for this scan. Only the first two elements (bscan, column) are used as the en-face reference; the third (the axial row) is carried through unused, since the en-face plane has already collapsed that axis.
  • roi_radius_mm: Radius in mm of a region of interest centred on fovea_centre, or None to skip it. Has no effect without a fovea reference.
  • max_lesions_per_group: Cap on lesions measured in detail per group; the largest are kept.
  • declared_num_bscans: The scan's own frame count, so trailing padding from the inference run's wider scans is not measured. See parse_bscan_masks.
  • angular_span_min_radius_mm: Radius inside which cells contribute no angle to aggregate_angular_span, or None for no floor. See LesionCalculationConfig.angular_span_min_radius_mm for why a trial has to choose one.

Returns The scan's per-group metrics.

Raises

  • Exception: Propagates any parsing or computation failure so the caller can record it against the file.