dissector.evaluation
====================

.. py:module:: dissector.evaluation


Functions
---------

.. autoapisummary::

   dissector.evaluation.compare_folders
   dissector.evaluation.compare_images
   dissector.evaluation.extract_boundary_2d
   dissector.evaluation.extract_boundary_3d
   dissector.evaluation.dilate_2d
   dissector.evaluation.dilate_3d
   dissector.evaluation.binary_cross_entropy
   dissector.evaluation.inter_slice_dice
   dissector.evaluation.boundary_iou_3d
   dissector.evaluation.boundary_iou_2d
   dissector.evaluation.single_row_viz
   dissector.evaluation.radarplot_single_dataset
   dissector.evaluation.side_by_side_comp
   dissector.evaluation.violin_compare


Module Contents
---------------

.. py:function:: compare_folders(folder_a: str, folder_b: str) -> bool

   Hash every file in two folders and assert they are identical.

   Compares file names and SHA-256 content. Raises AssertionError with a
   human-readable diff if anything differs. Returns True if identical.


.. py:function:: compare_images(path_a: str, path_b: str) -> bool

   Hash two image files and assert they are identical.

   Computes SHA-256 of each file and raises AssertionError if they differ.
   Returns True if the files are byte-for-byte identical.

   :param path_a: path to the first image file.
   :param path_b: path to the second image file.

   :returns: True if both files have the same SHA-256 hash.

   :raises AssertionError: if the hashes differ, with both paths in the message.


.. py:function:: extract_boundary_2d(mask: numpy.ndarray) -> numpy.ndarray

   Get boundary pixels of a binary mask on 2d image.


.. py:function:: extract_boundary_3d(mask: numpy.ndarray) -> numpy.ndarray

   Get boundary voxels of a 3D binary mask.


.. py:function:: dilate_2d(mask: numpy.ndarray, distance: int) -> numpy.ndarray

   Simple square dilation on a 2D mask using NumPy only.


.. py:function:: dilate_3d(mask: numpy.ndarray, distance: int) -> numpy.ndarray

   Simple dilation using NumPy only on a 3D matrix.


.. py:function:: binary_cross_entropy(y_true: numpy.ndarray, y_pred: numpy.ndarray, eps: float = 1e-07) -> float

   Binary cross entropy between two binary segmentation masks.

   Clips y_pred to [eps, 1-eps] so that exact 0/1 voxels don't produce log(0).


.. py:function:: inter_slice_dice(mask: numpy.ndarray) -> float

   Mean Dice between every pair of adjacent slices in a 3D binary mask.

   mask shape: (slices, H, W). Returns a value in [0, 1]; higher means
   smoother continuity across slices. Pairs where both slices are empty
   are skipped; a slice with no foreground paired against a non-empty one
   contributes 0.


.. py:function:: boundary_iou_3d(distance: int, mask_a: numpy.ndarray, mask_b: numpy.ndarray) -> float

   Boundary iou.

   From metric definition same as boundary_iou_2d but done in 3d.


.. py:function:: boundary_iou_2d(distance: int, mask_a: numpy.ndarray, mask_b: numpy.ndarray) -> float

   Extract boundary iou.

   According to
   formula from https://metrics-reloaded.dkfz.de/metric-library/
   boundary_intersection_over_union_local definition.


.. py:function:: single_row_viz(metrics: list, dataset: pandas.dataframe, row: int, title: str) -> None

   Graph a single image.

   This visualizes a single row of metrics on a radar plot. metrics is a
   list, dataset is the dataframe, row is the row number.


.. py:function:: radarplot_single_dataset(metrics: list[str], dataset: pandas.DataFrame, title: str, labels: list[str] | None = None) -> None

   Plot a radar chart for all rows in a dataset.

   If labels is None, metrics are used as axis labels.


.. py:function:: side_by_side_comp(datasets: list, metrics: list, titles: list, colors: list, labels: list) -> None

   Two radarplots.

   Warning, titles and colors must go in order of dataset1, dataset2


.. py:function:: violin_compare(csv_list_a: list[str], csv_list_b: list[str], label_a: str = 'Algorithm A', label_b: str = 'Algorithm B', metrics: list[str] | None = None, save_path: str | None = None) -> None

   Violin plots comparing two sets of segmentation results across CSV files.

   Each CSV list comes from one segmentation algorithm. Shared numeric columns
   (or an explicit metrics list) are plotted side by side as paired violins.