test_auto3dseg.py

# Copyright (c) MONAI Consortium
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#     http://www.apache.org/licenses/LICENSE-2.0
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from __future__ import annotations

import os
import tempfile
import unittest
from copy import deepcopy
from numbers import Number

import nibabel as nib
import numpy as np
import torch
from parameterized import parameterized

from monai.apps.auto3dseg import DataAnalyzer
from monai.auto3dseg import (
    Analyzer,
    FgImageStats,
    FgImageStatsSumm,
    FilenameStats,
    ImageStats,
    ImageStatsSumm,
    LabelStats,
    LabelStatsSumm,
    Operations,
    SampleOperations,
    SegSummarizer,
    SummaryOperations,
    datafold_read,
    verify_report_format,
)
from monai.bundle import ConfigParser
from monai.data import DataLoader, Dataset, create_test_image_2d, create_test_image_3d
from monai.data.meta_tensor import MetaTensor
from monai.data.utils import no_collation
from monai.transforms import (
    Compose,
    EnsureChannelFirstd,
    EnsureTyped,
    Lambdad,
    LoadImaged,
    Orientationd,
    SqueezeDimd,
    ToDeviced,
)
from monai.utils.enums import DataStatsKeys, LabelStatsKeys
from tests.test_utils import skip_if_no_cuda

device = "cpu"
n_workers = 2

sim_datalist = {
    "testing": [{"image": "val_001.fake.nii.gz"}, {"image": "val_002.fake.nii.gz"}],
    "training": [
        {"fold": 0, "image": "tr_image_001.fake.nii.gz", "label": "tr_label_001.fake.nii.gz"},
        {"fold": 0, "image": "tr_image_002.fake.nii.gz", "label": "tr_label_002.fake.nii.gz"},
        {"fold": 1, "image": "tr_image_001.fake.nii.gz", "label": "tr_label_001.fake.nii.gz"},
        {"fold": 1, "image": "tr_image_004.fake.nii.gz", "label": "tr_label_004.fake.nii.gz"},
    ],
}

SIM_CPU_TEST_CASES = [
    [{"sim_dim": (32, 32, 32), "label_key": "label"}],
    [{"sim_dim": (32, 32, 32, 2), "label_key": "label"}],
    [{"sim_dim": (32, 32, 32), "label_key": None}],
    [{"sim_dim": (32, 32, 32), "label_key": "None"}],
]

SIM_GPU_TEST_CASES = [[{"sim_dim": (32, 32, 32), "label_key": "label"}], [{"sim_dim": (32, 32, 32), "label_key": None}]]

LABEL_STATS_DEVICE_TEST_CASES = [
    [{"image_device": "cpu", "label_device": "cpu", "image_meta": False}],
    [{"image_device": "cuda", "label_device": "cuda", "image_meta": True}],
    [{"image_device": "cpu", "label_device": "cuda", "image_meta": True}],
    [{"image_device": "cuda", "label_device": "cpu", "image_meta": False}],
]


def create_sim_data(dataroot: str, sim_datalist: dict, sim_dim: tuple, image_only: bool = False, **kwargs) -> None:
    """
    Create simulated data using create_test_image_3d.

    Args:
        dataroot: data directory path that hosts the "nii.gz" image files.
        sim_datalist: a list of data to create.
        sim_dim: the image sizes, for examples: a tuple of (64, 64, 64) for 3d, or (128, 128) for 2d
    """
    if not os.path.isdir(dataroot):
        os.makedirs(dataroot)

    # Generate a fake dataset
    for d in sim_datalist["testing"] + sim_datalist["training"]:
        if len(sim_dim) == 2:  # 2D image
            im, seg = create_test_image_2d(sim_dim[0], sim_dim[1], **kwargs)
        elif len(sim_dim) == 3:  # 3D image
            im, seg = create_test_image_3d(sim_dim[0], sim_dim[1], sim_dim[2], **kwargs)
        elif len(sim_dim) == 4:  # multi-modality 3D image
            im_list = []
            seg_list = []
            for _ in range(sim_dim[3]):
                im_3d, seg_3d = create_test_image_3d(sim_dim[0], sim_dim[1], sim_dim[2], **kwargs)
                im_list.append(im_3d[..., np.newaxis])
                seg_list.append(seg_3d[..., np.newaxis])
            im = np.concatenate(im_list, axis=3)
            seg = np.concatenate(seg_list, axis=3)
        else:
            raise ValueError(f"Invalid argument input. sim_dim has f{len(sim_dim)} values. 2-4 values are expected.")
        nib_image = nib.Nifti1Image(im, affine=np.eye(4))
        image_fpath = os.path.join(dataroot, d["image"])
        nib.save(nib_image, image_fpath)

        if not image_only and "label" in d:
            nib_image = nib.Nifti1Image(seg, affine=np.eye(4))
            label_fpath = os.path.join(dataroot, d["label"])
            nib.save(nib_image, label_fpath)


class TestOperations(Operations):
    """
    Test example for user operation
    """

    __test__ = False  # indicate to pytest that this class is not intended for collection

    def __init__(self) -> None:
        self.data = {"max": np.max, "mean": np.mean, "min": np.min}


class TestAnalyzer(Analyzer):
    """
    Test example for a simple Analyzer
    """

    __test__ = False  # indicate to pytest that this class is not intended for collection

    def __init__(self, key, report_format, stats_name="test"):
        self.key = key
        super().__init__(stats_name, report_format)

    def __call__(self, data):
        d = dict(data)
        report = deepcopy(self.get_report_format())
        report["stats"] = self.ops["stats"].evaluate(d[self.key])
        d[self.stats_name] = report
        return d


class TestImageAnalyzer(Analyzer):
    """
    Test example for a simple Analyzer
    """

    __test__ = False  # indicate to pytest that this class is not intended for collection

    def __init__(self, image_key="image", stats_name="test_image"):
        self.image_key = image_key
        report_format = {"test_stats": None}

        super().__init__(stats_name, report_format)
        self.update_ops("test_stats", TestOperations())

    def __call__(self, data):
        d = dict(data)
        report = deepcopy(self.get_report_format())
        report["test_stats"] = self.ops["test_stats"].evaluate(d[self.image_key])
        d[self.stats_name] = report
        return d


class TestDataAnalyzer(unittest.TestCase):
    """Integration tests for the auto3dseg analyzer pipeline."""

    def setUp(self):
        """Create temporary directory and write simulated datalist JSON file."""
        self.test_dir = tempfile.TemporaryDirectory()
        work_dir = self.test_dir.name
        self.dataroot_dir = os.path.join(work_dir, "sim_dataroot")
        self.datalist_file = os.path.join(work_dir, "sim_datalist.json")
        self.datastat_file = os.path.join(work_dir, "datastats.yml")
        ConfigParser.export_config_file(sim_datalist, self.datalist_file)

    @parameterized.expand(SIM_CPU_TEST_CASES)
    def test_data_analyzer_cpu(self, input_params):
        """Verify DataAnalyzer produces per-case stats on CPU across dim/label combinations."""
        sim_dim = input_params["sim_dim"]
        label_key = input_params["label_key"]
        image_only = not bool(label_key)
        rmax = max(int(sim_dim[0] / 4), 1)
        create_sim_data(
            self.dataroot_dir, sim_datalist, sim_dim, image_only=image_only, rad_max=rmax, rad_min=1, num_seg_classes=1
        )

        analyser = DataAnalyzer(
            self.datalist_file, self.dataroot_dir, output_path=self.datastat_file, label_key=label_key, device=device
        )
        datastat = analyser.get_all_case_stats()

        assert len(datastat["stats_by_cases"]) == len(sim_datalist["training"])

    def test_data_analyzer_histogram(self):
        """Verify DataAnalyzer runs in histogram_only mode with no label key."""
        create_sim_data(
            self.dataroot_dir, sim_datalist, [32] * 3, image_only=True, rad_max=8, rad_min=1, num_seg_classes=1
        )
        analyser = DataAnalyzer(
            self.datalist_file,
            self.dataroot_dir,
            output_path=self.datastat_file,
            label_key=None,
            device=device,
            histogram_only=True,
        )
        datastat = analyser.get_all_case_stats()
        assert len(datastat["stats_by_cases"]) == len(sim_datalist["training"])

    @parameterized.expand(SIM_GPU_TEST_CASES)
    @skip_if_no_cuda
    def test_data_analyzer_gpu(self, input_params):
        """Verify DataAnalyzer produces per-case stats on GPU (skipped if CUDA unavailable)."""
        sim_dim = input_params["sim_dim"]
        label_key = input_params["label_key"]
        image_only = not bool(label_key)
        rmax = max(int(sim_dim[0] / 4), 1)
        create_sim_data(
            self.dataroot_dir, sim_datalist, sim_dim, image_only=image_only, rad_max=rmax, rad_min=1, num_seg_classes=1
        )
        analyser = DataAnalyzer(
            self.datalist_file, self.dataroot_dir, output_path=self.datastat_file, label_key=label_key, device="cuda"
        )
        datastat = analyser.get_all_case_stats()

        assert len(datastat["stats_by_cases"]) == len(sim_datalist["training"])

    def test_basic_operation_class(self):
        """Verify Operations.evaluate returns correct stat keys and shapes with and without axis."""
        op = TestOperations()
        test_data = np.random.rand(10, 10).astype(np.float64)
        test_ret_1 = op.evaluate(test_data)
        test_ret_2 = op.evaluate(test_data, axis=0)
        assert isinstance(test_ret_1, dict) and isinstance(test_ret_2, dict)
        assert ("max" in test_ret_1) and ("max" in test_ret_2)
        assert ("mean" in test_ret_1) and ("mean" in test_ret_2)
        assert ("min" in test_ret_1) and ("min" in test_ret_2)
        assert isinstance(test_ret_1["max"], np.float64)
        assert isinstance(test_ret_2["max"], np.ndarray)
        assert test_ret_1["max"].ndim == 0
        assert test_ret_2["max"].ndim == 1

    def test_sample_operations(self):
        """Verify SampleOperations works with both numpy arrays and MetaTensors."""
        op = SampleOperations()
        test_data_np = np.random.rand(10, 10).astype(np.float64)
        test_data_mt = MetaTensor(test_data_np, device=device)
        test_ret_np = op.evaluate(test_data_np)
        test_ret_mt = op.evaluate(test_data_mt)
        assert isinstance(test_ret_np["max"], Number)
        assert isinstance(test_ret_np["percentile"], list)
        assert isinstance(test_ret_mt["max"], Number)
        assert isinstance(test_ret_mt["percentile"], list)

        op.update({"sum": np.sum})
        test_ret_np = op.evaluate(test_data_np)
        assert "sum" in test_ret_np

    def test_summary_operations(self):
        """Verify SummaryOperations reduces a stat dict to scalar summary values."""
        op = SummaryOperations()
        test_dict = {"min": [0, 1, 2, 3], "max": [2, 3, 4, 5], "mean": [1, 2, 3, 4], "sum": [2, 4, 6, 8]}
        test_ret = op.evaluate(test_dict)
        assert isinstance(test_ret["max"], Number)
        assert isinstance(test_ret["min"], Number)

        op.update({"sum": np.sum})
        test_ret = op.evaluate(test_dict)
        assert "sum" in test_ret
        assert isinstance(test_ret["sum"], Number)

    def test_basic_analyzer_class(self):
        """Verify a custom Analyzer subclass computes and stores stats in the output dict."""
        test_data = {}
        test_data["image_test"] = np.random.rand(10, 10)
        report_format = {"stats": None}
        user_analyzer = TestAnalyzer("image_test", report_format)
        user_analyzer.update_ops("stats", TestOperations())
        result = user_analyzer(test_data)
        assert result["test"]["stats"]["max"] == np.max(test_data["image_test"])
        assert result["test"]["stats"]["min"] == np.min(test_data["image_test"])
        assert result["test"]["stats"]["mean"] == np.mean(test_data["image_test"])

    def test_transform_analyzer_class(self):
        """Verify a custom Analyzer integrates correctly as a step in a Compose transform."""
        transform = Compose([LoadImaged(keys=["image"]), TestImageAnalyzer(image_key="image")])
        create_sim_data(self.dataroot_dir, sim_datalist, (32, 32, 32), rad_max=8, rad_min=1, num_seg_classes=1)
        files, _ = datafold_read(sim_datalist, self.dataroot_dir, fold=-1)
        ds = Dataset(data=files)
        self.dataset = DataLoader(ds, batch_size=1, shuffle=False, num_workers=0, collate_fn=no_collation)
        for batch_data in self.dataset:
            d = transform(batch_data[0])
            assert "test_image" in d
            assert "test_stats" in d["test_image"]
            assert "max" in d["test_image"]["test_stats"]
            assert "min" in d["test_image"]["test_stats"]
            assert "mean" in d["test_image"]["test_stats"]

    def test_image_stats_case_analyzer(self):
        """Verify ImageStats produces a report matching the expected format for 3-D images."""
        analyzer = ImageStats(image_key="image")
        transform = Compose(
            [
                LoadImaged(keys=["image"]),
                EnsureChannelFirstd(keys=["image"]),  # this creates label to be (1,H,W,D)
                ToDeviced(keys=["image"], device=device, non_blocking=True),
                Orientationd(keys=["image"], axcodes="RAS"),
                EnsureTyped(keys=["image"], data_type="tensor"),
                analyzer,
            ]
        )
        create_sim_data(self.dataroot_dir, sim_datalist, (32, 32, 32), rad_max=8, rad_min=1, num_seg_classes=1)
        files, _ = datafold_read(sim_datalist, self.dataroot_dir, fold=-1)
        ds = Dataset(data=files)
        self.dataset = DataLoader(ds, batch_size=1, shuffle=False, num_workers=n_workers, collate_fn=no_collation)
        for batch_data in self.dataset:
            d = transform(batch_data[0])
            report_format = analyzer.get_report_format()
            assert verify_report_format(d["image_stats"], report_format)

    def test_foreground_image_stats_cases_analyzer(self):
        """Verify FgImageStats produces a valid foreground stats report."""
        analyzer = FgImageStats(image_key="image", label_key="label")
        transform_list = [
            LoadImaged(keys=["image", "label"]),
            EnsureChannelFirstd(keys=["image", "label"]),  # this creates label to be (1,H,W,D)
            ToDeviced(keys=["image", "label"], device=device, non_blocking=True),
            Orientationd(keys=["image", "label"], axcodes="RAS"),
            EnsureTyped(keys=["image", "label"], data_type="tensor"),
            Lambdad(keys=["label"], func=lambda x: torch.argmax(x, dim=0, keepdim=True) if x.shape[0] > 1 else x),
            SqueezeDimd(keys=["label"], dim=0),
            analyzer,
        ]
        transform = Compose(transform_list)
        create_sim_data(self.dataroot_dir, sim_datalist, (32, 32, 32), rad_max=8, rad_min=1, num_seg_classes=1)
        files, _ = datafold_read(sim_datalist, self.dataroot_dir, fold=-1)
        ds = Dataset(data=files)
        self.dataset = DataLoader(ds, batch_size=1, shuffle=False, num_workers=n_workers, collate_fn=no_collation)
        for batch_data in self.dataset:
            d = transform(batch_data[0])
            report_format = analyzer.get_report_format()
            assert verify_report_format(d["image_foreground_stats"], report_format)

    def test_label_stats_case_analyzer(self):
        """Verify LabelStats produces a valid report including per-label statistics."""
        analyzer = LabelStats(image_key="image", label_key="label")
        transform = Compose(
            [
                LoadImaged(keys=["image", "label"]),
                EnsureChannelFirstd(keys=["image", "label"]),  # this creates label to be (1,H,W,D)
                ToDeviced(keys=["image", "label"], device=device, non_blocking=True),
                Orientationd(keys=["image", "label"], axcodes="RAS"),
                EnsureTyped(keys=["image", "label"], data_type="tensor"),
                Lambdad(keys=["label"], func=lambda x: torch.argmax(x, dim=0, keepdim=True) if x.shape[0] > 1 else x),
                SqueezeDimd(keys=["label"], dim=0),
                analyzer,
            ]
        )
        create_sim_data(self.dataroot_dir, sim_datalist, (32, 32, 32), rad_max=8, rad_min=1, num_seg_classes=1)
        files, _ = datafold_read(sim_datalist, self.dataroot_dir, fold=-1)
        ds = Dataset(data=files)
        self.dataset = DataLoader(ds, batch_size=1, shuffle=False, num_workers=n_workers, collate_fn=no_collation)
        for batch_data in self.dataset:
            d = transform(batch_data[0])
            report_format = analyzer.get_report_format()
            assert verify_report_format(d["label_stats"], report_format)

    @parameterized.expand(LABEL_STATS_DEVICE_TEST_CASES)
    def test_label_stats_mixed_device_analyzer(self, input_params):
        """Verify LabelStats handles tensors split across CPU and CUDA devices."""
        image_device = torch.device(input_params["image_device"])
        label_device = torch.device(input_params["label_device"])

        if (image_device.type == "cuda" or label_device.type == "cuda") and not torch.cuda.is_available():
            self.skipTest("CUDA is not available for mixed-device LabelStats tests.")

        analyzer = LabelStats(image_key="image", label_key="label")

        image_tensor = torch.tensor(
            [
                [[[1.0, 2.0], [3.0, 4.0]], [[5.0, 6.0], [7.0, 8.0]]],
                [[[11.0, 12.0], [13.0, 14.0]], [[15.0, 16.0], [17.0, 18.0]]],
            ],
            dtype=torch.float32,
        ).to(image_device)
        label_tensor = torch.tensor([[[0, 1], [1, 0]], [[0, 1], [0, 1]]], dtype=torch.int64).to(label_device)

        if input_params["image_meta"]:
            image_tensor = MetaTensor(image_tensor)
        label_tensor = MetaTensor(label_tensor)

        result = analyzer({"image": image_tensor, "label": label_tensor})
        report = result["label_stats"]

        # Verify report format and computation succeeded despite mixed/unified devices
        assert verify_report_format(report, analyzer.get_report_format())
        assert report[LabelStatsKeys.LABEL_UID] == [0, 1]

        label_stats = report[LabelStatsKeys.LABEL]
        self.assertAlmostEqual(label_stats[0][LabelStatsKeys.PIXEL_PCT], 0.5)
        self.assertAlmostEqual(label_stats[1][LabelStatsKeys.PIXEL_PCT], 0.5)

        label0_intensity = label_stats[0][LabelStatsKeys.IMAGE_INTST]
        label1_intensity = label_stats[1][LabelStatsKeys.IMAGE_INTST]
        self.assertAlmostEqual(label0_intensity[0]["mean"], 4.25)
        self.assertAlmostEqual(label1_intensity[0]["mean"], 4.75)
        self.assertAlmostEqual(label0_intensity[1]["mean"], 14.25)
        self.assertAlmostEqual(label1_intensity[1]["mean"], 14.75)

        foreground_stats = report[LabelStatsKeys.IMAGE_INTST]
        self.assertAlmostEqual(foreground_stats[0]["mean"], 4.75)
        self.assertAlmostEqual(foreground_stats[1]["mean"], 14.75)

    def test_filename_case_analyzer(self):
        """Verify FilenameStats records both image and label paths in the output dict."""
        analyzer_image = FilenameStats("image", DataStatsKeys.BY_CASE_IMAGE_PATH)
        analyzer_label = FilenameStats("label", DataStatsKeys.BY_CASE_IMAGE_PATH)
        transform_list = [LoadImaged(keys=["image", "label"]), analyzer_image, analyzer_label]
        transform = Compose(transform_list)
        create_sim_data(self.dataroot_dir, sim_datalist, (32, 32, 32), rad_max=8, rad_min=1, num_seg_classes=1)
        files, _ = datafold_read(sim_datalist, self.dataroot_dir, fold=-1)
        ds = Dataset(data=files)
        self.dataset = DataLoader(ds, batch_size=1, shuffle=False, num_workers=n_workers, collate_fn=no_collation)
        for batch_data in self.dataset:
            d = transform(batch_data[0])
            assert DataStatsKeys.BY_CASE_IMAGE_PATH in d

    def test_filename_case_analyzer_image_only(self):
        """Verify FilenameStats handles image-only input and stores 'None' for the label path."""
        analyzer_image = FilenameStats("image", DataStatsKeys.BY_CASE_IMAGE_PATH)
        analyzer_label = FilenameStats(None, DataStatsKeys.BY_CASE_IMAGE_PATH)
        transform_list = [LoadImaged(keys=["image"]), analyzer_image, analyzer_label]
        transform = Compose(transform_list)
        create_sim_data(self.dataroot_dir, sim_datalist, (32, 32, 32), rad_max=8, rad_min=1, num_seg_classes=1)
        files, _ = datafold_read(sim_datalist, self.dataroot_dir, fold=-1)
        ds = Dataset(data=files)
        self.dataset = DataLoader(ds, batch_size=1, shuffle=False, num_workers=n_workers, collate_fn=no_collation)
        for batch_data in self.dataset:
            d = transform(batch_data[0])
            assert DataStatsKeys.BY_CASE_IMAGE_PATH in d
            assert d[DataStatsKeys.BY_CASE_IMAGE_PATH] == "None"

    def test_image_stats_summary_analyzer(self):
        """Verify ImageStatsSumm correctly aggregates per-case image stats."""
        summary_analyzer = ImageStatsSumm("image_stats")

        transform_list = [
            LoadImaged(keys=["image"]),
            EnsureChannelFirstd(keys=["image"]),  # this creates label to be (1,H,W,D)
            ToDeviced(keys=["image"], device=device, non_blocking=True),
            Orientationd(keys=["image"], axcodes="RAS"),
            EnsureTyped(keys=["image"], data_type="tensor"),
            ImageStats(image_key="image"),
        ]
        transform = Compose(transform_list)
        create_sim_data(self.dataroot_dir, sim_datalist, (32, 32, 32), rad_max=8, rad_min=1, num_seg_classes=1)
        files, _ = datafold_read(sim_datalist, self.dataroot_dir, fold=-1)
        ds = Dataset(data=files)
        self.dataset = DataLoader(ds, batch_size=1, shuffle=False, num_workers=n_workers, collate_fn=no_collation)
        stats = []
        for batch_data in self.dataset:
            stats.append(transform(batch_data[0]))
        summary_report = summary_analyzer(stats)
        report_format = summary_analyzer.get_report_format()
        assert verify_report_format(summary_report, report_format)

    def test_fg_image_stats_summary_analyzer(self):
        """Verify FgImageStatsSumm correctly aggregates per-case foreground stats."""
        summary_analyzer = FgImageStatsSumm("image_foreground_stats")

        transform_list = [
            LoadImaged(keys=["image", "label"]),
            EnsureChannelFirstd(keys=["image", "label"]),  # this creates label to be (1,H,W,D)
            ToDeviced(keys=["image", "label"], device=device, non_blocking=True),
            Orientationd(keys=["image", "label"], axcodes="RAS"),
            EnsureTyped(keys=["image", "label"], data_type="tensor"),
            Lambdad(keys="label", func=lambda x: torch.argmax(x, dim=0, keepdim=True) if x.shape[0] > 1 else x),
            SqueezeDimd(keys=["label"], dim=0),
            FgImageStats(image_key="image", label_key="label"),
        ]
        transform = Compose(transform_list)
        create_sim_data(self.dataroot_dir, sim_datalist, (32, 32, 32), rad_max=8, rad_min=1, num_seg_classes=1)
        files, _ = datafold_read(sim_datalist, self.dataroot_dir, fold=-1)
        ds = Dataset(data=files)
        self.dataset = DataLoader(ds, batch_size=1, shuffle=False, num_workers=n_workers, collate_fn=no_collation)
        stats = []
        for batch_data in self.dataset:
            stats.append(transform(batch_data[0]))
        summary_report = summary_analyzer(stats)
        report_format = summary_analyzer.get_report_format()
        assert verify_report_format(summary_report, report_format)

    def test_label_stats_summary_analyzer(self):
        """Verify LabelStatsSumm correctly aggregates per-case label stats."""
        summary_analyzer = LabelStatsSumm("label_stats")

        transform_list = [
            LoadImaged(keys=["image", "label"]),
            EnsureChannelFirstd(keys=["image", "label"]),  # this creates label to be (1,H,W,D)
            ToDeviced(keys=["image", "label"], device=device, non_blocking=True),
            Orientationd(keys=["image", "label"], axcodes="RAS"),
            EnsureTyped(keys=["image", "label"], data_type="tensor"),
            Lambdad(keys="label", func=lambda x: torch.argmax(x, dim=0, keepdim=True) if x.shape[0] > 1 else x),
            SqueezeDimd(keys=["label"], dim=0),
            LabelStats(image_key="image", label_key="label"),
        ]
        transform = Compose(transform_list)
        create_sim_data(self.dataroot_dir, sim_datalist, (32, 32, 32), rad_max=8, rad_min=1, num_seg_classes=1)
        files, _ = datafold_read(sim_datalist, self.dataroot_dir, fold=-1)
        ds = Dataset(data=files)
        self.dataset = DataLoader(ds, batch_size=1, shuffle=False, num_workers=n_workers, collate_fn=no_collation)
        stats = []
        for batch_data in self.dataset:
            stats.append(transform(batch_data[0]))
        summary_report = summary_analyzer(stats)
        report_format = summary_analyzer.get_report_format()
        assert verify_report_format(summary_report, report_format)

    def test_seg_summarizer(self):
        """Verify SegSummarizer produces a summary with image, foreground, and label stat keys."""
        summarizer = SegSummarizer("image", "label")
        keys = ["image", "label"]
        transform_list = [
            LoadImaged(keys=keys),
            EnsureChannelFirstd(keys=keys),  # this creates label to be (1,H,W,D)
            ToDeviced(keys=keys, device=device, non_blocking=True),
            Orientationd(keys=keys, axcodes="RAS"),
            EnsureTyped(keys=keys, data_type="tensor"),
            Lambdad(keys="label", func=lambda x: torch.argmax(x, dim=0, keepdim=True) if x.shape[0] > 1 else x),
            SqueezeDimd(keys=["label"], dim=0),
            summarizer,
        ]
        transform = Compose(transform_list)
        create_sim_data(self.dataroot_dir, sim_datalist, (32, 32, 32), rad_max=8, rad_min=1, num_seg_classes=1)
        files, _ = datafold_read(sim_datalist, self.dataroot_dir, fold=-1)
        ds = Dataset(data=files)
        self.dataset = DataLoader(ds, batch_size=1, shuffle=False, num_workers=n_workers, collate_fn=no_collation)
        stats = []
        for batch_data in self.dataset:
            d = transform(batch_data[0])
            stats.append(d)
        report = summarizer.summarize(stats)
        assert str(DataStatsKeys.IMAGE_STATS) in report
        assert str(DataStatsKeys.FG_IMAGE_STATS) in report
        assert str(DataStatsKeys.LABEL_STATS) in report

    def test_image_stats_precomputed_nda_croppeds(self):
        """Verify ImageStats handles pre-populated nda_croppeds without crashing.

        Previously raised UnboundLocalError because nda_croppeds was only assigned
        inside the ``if "nda_croppeds" not in d`` branch but used unconditionally.
        """
        analyzer = ImageStats(image_key="image")
        image = torch.rand(1, 10, 10, 10)
        precomputed = [np.random.rand(8, 8, 8)]  # simulated pre-cropped foreground
        data = {"image": MetaTensor(image), "nda_croppeds": precomputed}
        result = analyzer(data)
        assert "image_stats" in result
        assert verify_report_format(result["image_stats"], analyzer.get_report_format())

    def test_analyzer_grad_state_restored_after_call(self):
        """Verify ImageStats restores torch grad-enabled state on both normal and disabled entry.

        Checks that the try/finally guard correctly restores the state regardless of
        whether grad was enabled or disabled before the call.
        """
        analyzer = ImageStats(image_key="image")
        image = torch.rand(1, 10, 10, 10)
        data = {"image": MetaTensor(image)}

        # grad enabled before call → must still be enabled after
        torch.set_grad_enabled(True)
        analyzer(data)
        assert torch.is_grad_enabled(), "grad state was not restored after ImageStats call"

        # grad disabled before call → must still be disabled after
        torch.set_grad_enabled(False)
        try:
            analyzer(data)
            assert not torch.is_grad_enabled(), "grad state was not restored after ImageStats call"
        finally:
            torch.set_grad_enabled(True)  # always restore for subsequent tests

    def tearDown(self) -> None:
        """Remove the temporary test directory."""
        self.test_dir.cleanup()


if __name__ == "__main__":
    unittest.main()