Skip to main content

mask_groups

A named group of segmentation labels and how its mask is combined.

Shared vocabulary rather than step-local config: lesion_calculation measures groups, and any later per-lesion morphometry measures the SAME groups, so the definition of "what counts as GA in this trial" lives in one place.

Classes​

LesionConnectivity​

class LesionConnectivity(*args, **kwds):

Whether a lesion is one component on the en-face plane or in the volume.

EN_FACE matches what a clinician counts looking at the projection, and is right for drusen, which sit in a thin band above Bruch's membrane. VOLUME separates biomarkers that genuinely stack axially — two fluid cysts in the same column are two cysts.

StrEnum for the same reason as MaskCombineMode.

Ancestors​

Variables​

  • static EN_FACE
  • static VOLUME

MaskCombineMode​

class MaskCombineMode(*args, **kwds):

How a group's include labels are combined into one mask.

intersection is the GA definition: a column counts only where EVERY include label is present (atrophy is evidenced by co-located layers). union is the drusen definition: hard, soft and confluent drusen are alternative subtypes, so a column counts where ANY of them is present. Summing the three per-label areas instead would double-count any column carrying two subtypes.

StrEnum, so a member is its own YAML/JSON representation: it binds from the plain string a template writes, and model_dump(mode="json") emits that same string — which is what flows/dag/hashing.py hashes into task_hash, so the vocabulary can gain a spelling without rotating a cache partition.

Ancestors​

Variables​

  • static INTERSECTION
  • static UNION

MaskGroupSpec​

class MaskGroupSpec(**data: Any):

One named group of segmentation labels, and how to combine them.

Attributes

  • include: Labels whose masks are combined per combine. An empty list includes nothing (matching the GA parser), never everything.
  • exclude: Labels whose presence removes a cell, combined as a union and applied after include. An empty list excludes nothing.
  • combine: intersection (GA) or union (drusen).
  • na_bscans: How to treat B-scans the model produced no output for. retain keeps them as zero rows; drop removes them before measuring, compacting the B-scan axis.
  • lesion_connectivity: Whether a lesion is one connected component on the en-face plane or in the full volume.
  • measure_lesions: Whether to measure each lesion individually — its own area, volume, diameters and fovea-relative geometry — instead of only the group totals. Off by default because labelling and measuring every lesion is expensive relative to an area sum.

Raises

  • ValueError: If a label appears in both include and exclude.

Create a new model by parsing and validating input data from keyword arguments.

Raises [ValidationError][pydantic_core.ValidationError] if the input data cannot be validated to form a valid model.

self is explicitly positional-only to allow self as a field name.

Variables​

  • static exclude : list[str]
  • static include : list[str]
  • static measure_lesions : bool
  • static model_config

NaBscanMode​

class NaBscanMode(*args, **kwds):

How a group treats B-scans the model produced no output for.

RETAIN measures them as empty at their true index, which keeps two lesions either side of a skipped frame distinct. DROP removes them before measuring, compacting the B-scan axis — the GA path's behaviour, and what a group must use to reproduce GA's published numbers. Under DROP, max_area_bscan_index indexes the compacted axis.

Neither mode sees the trailing padding a shorter scan carries: the inference cache stores one key per column of a run-wide DataFrame, so a file with fewer frames than the widest in its run is padded with nulls. parse_bscan_masks truncates that padding for every group before either mode applies, so distance_from_image_centre is centred on the scan's own frames rather than the run's widest.

StrEnum for the same reason as MaskCombineMode.

Ancestors​

Variables​

  • static DROP
  • static RETAIN