Skip to content

Commit 7ab30b5

Browse files
Merge branch 'dev' into improve-affine-docs-7092
2 parents c70bb0b + 8d39519 commit 7ab30b5

File tree

13 files changed

+585
-25
lines changed

13 files changed

+585
-25
lines changed

docs/source/losses.rst

Lines changed: 5 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -98,6 +98,11 @@ Segmentation Losses
9898
.. autoclass:: NACLLoss
9999
:members:
100100

101+
`MCCLoss`
102+
~~~~~~~~~
103+
.. autoclass:: MCCLoss
104+
:members:
105+
101106
Registration Losses
102107
-------------------
103108

monai/apps/auto3dseg/auto_runner.py

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -229,7 +229,7 @@ def __init__(
229229
input = os.path.join(os.path.abspath(work_dir), "input.yaml")
230230
logger.info(f"Input config is not provided, using the default {input}")
231231

232-
self.data_src_cfg = dict()
232+
self.data_src_cfg = {}
233233
if isinstance(input, dict):
234234
self.data_src_cfg = input
235235
elif isinstance(input, str) and os.path.isfile(input):

monai/apps/reconstruction/transforms/dictionary.py

Lines changed: 34 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -20,6 +20,7 @@
2020
from monai.apps.reconstruction.transforms.array import EquispacedKspaceMask, RandomKspaceMask
2121
from monai.config import DtypeLike, KeysCollection
2222
from monai.config.type_definitions import NdarrayOrTensor
23+
from monai.data.meta_tensor import MetaTensor
2324
from monai.transforms import InvertibleTransform
2425
from monai.transforms.croppad.array import SpatialCrop
2526
from monai.transforms.intensity.array import NormalizeIntensity
@@ -33,15 +34,36 @@ class ExtractDataKeyFromMetaKeyd(MapTransform):
3334
Moves keys from meta to data. It is useful when a dataset of paired samples
3435
is loaded and certain keys should be moved from meta to data.
3536
37+
This transform supports two modes:
38+
39+
1. When ``meta_key`` references a metadata dictionary in the data (e.g., when
40+
``image_only=False`` was used with ``LoadImaged``), the requested keys are
41+
extracted directly from that dictionary.
42+
43+
2. When ``meta_key`` references a ``MetaTensor`` in the data (e.g., when
44+
``image_only=True`` was used with ``LoadImaged``), the requested keys are
45+
extracted from its ``.meta`` attribute.
46+
3647
Args:
3748
keys: keys to be transferred from meta to data
38-
meta_key: the meta key where all the meta-data is stored
49+
meta_key: the key in the data dictionary where the metadata source is
50+
stored. This can be either a metadata dictionary or a ``MetaTensor``.
3951
allow_missing_keys: don't raise exception if key is missing
4052
4153
Example:
4254
When the fastMRI dataset is loaded, "kspace" is stored in the data dictionary,
4355
but the ground-truth image with the key "reconstruction_rss" is stored in the meta data.
4456
In this case, ExtractDataKeyFromMetaKeyd moves "reconstruction_rss" to data.
57+
58+
When ``LoadImaged`` is used with ``image_only=True`` (the default), the loaded
59+
data is a ``MetaTensor`` with metadata accessible via ``.meta``. In this case,
60+
set ``meta_key`` to the key of the ``MetaTensor`` itself::
61+
62+
li = LoadImaged(keys="image") # image_only=True by default
63+
dat = li({"image": "image.nii"})
64+
e = ExtractDataKeyFromMetaKeyd("filename_or_obj", meta_key="image")
65+
dat = e(dat)
66+
assert dat["image"].meta["filename_or_obj"] == dat["filename_or_obj"]
4567
"""
4668

4769
def __init__(self, keys: KeysCollection, meta_key: str, allow_missing_keys: bool = False) -> None:
@@ -58,9 +80,18 @@ def __call__(self, data: Mapping[Hashable, NdarrayOrTensor]) -> dict[Hashable, T
5880
the new data dictionary
5981
"""
6082
d = dict(data)
83+
meta_obj = d[self.meta_key]
84+
85+
# If meta_key references a MetaTensor, extract from its .meta attribute;
86+
# otherwise treat it as a metadata dictionary directly.
87+
if isinstance(meta_obj, MetaTensor):
88+
meta_dict: dict = meta_obj.meta
89+
else:
90+
meta_dict = dict(meta_obj)
91+
6192
for key in self.keys:
62-
if key in d[self.meta_key]:
63-
d[key] = d[self.meta_key][key] # type: ignore
93+
if key in meta_dict:
94+
d[key] = meta_dict[key] # type: ignore
6495
elif not self.allow_missing_keys:
6596
raise KeyError(
6697
f"Key `{key}` of transform `{self.__class__.__name__}` was missing in the meta data"

monai/auto3dseg/analyzer.py

Lines changed: 22 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -15,7 +15,7 @@
1515
from abc import ABC, abstractmethod
1616
from collections.abc import Hashable, Mapping
1717
from copy import deepcopy
18-
from typing import Any
18+
from typing import Any, cast
1919

2020
import numpy as np
2121
import torch
@@ -468,21 +468,35 @@ def __call__(self, data: Mapping[Hashable, MetaTensor]) -> dict[Hashable, MetaTe
468468
"""
469469
d: dict[Hashable, MetaTensor] = dict(data)
470470
start = time.time()
471-
if isinstance(d[self.image_key], (torch.Tensor, MetaTensor)) and d[self.image_key].device.type == "cuda":
472-
using_cuda = True
473-
else:
474-
using_cuda = False
471+
image_tensor = d[self.image_key]
472+
label_tensor = d[self.label_key]
473+
# Check if either tensor is on CUDA to determine if we should move both to CUDA for processing
474+
using_cuda = any(
475+
isinstance(t, (torch.Tensor, MetaTensor)) and t.device.type == "cuda" for t in (image_tensor, label_tensor)
476+
)
475477
restore_grad_state = torch.is_grad_enabled()
476478
torch.set_grad_enabled(False)
477479

478-
ndas: list[MetaTensor] = [d[self.image_key][i] for i in range(d[self.image_key].shape[0])] # type: ignore
479-
ndas_label: MetaTensor = d[self.label_key].astype(torch.int16) # (H,W,D)
480+
if isinstance(image_tensor, (MetaTensor, torch.Tensor)) and isinstance(
481+
label_tensor, (MetaTensor, torch.Tensor)
482+
):
483+
if label_tensor.device != image_tensor.device:
484+
if using_cuda:
485+
# Move both tensors to CUDA when mixing devices
486+
cuda_device = image_tensor.device if image_tensor.device.type == "cuda" else label_tensor.device
487+
image_tensor = cast(MetaTensor, image_tensor.to(cuda_device))
488+
label_tensor = cast(MetaTensor, label_tensor.to(cuda_device))
489+
else:
490+
label_tensor = cast(MetaTensor, label_tensor.to(image_tensor.device))
491+
492+
ndas: list[MetaTensor] = [image_tensor[i] for i in range(image_tensor.shape[0])] # type: ignore
493+
ndas_label: MetaTensor = label_tensor.astype(torch.int16) # (H,W,D)
480494

481495
if ndas_label.shape != ndas[0].shape:
482496
raise ValueError(f"Label shape {ndas_label.shape} is different from image shape {ndas[0].shape}")
483497

484498
nda_foregrounds: list[torch.Tensor] = [get_foreground_label(nda, ndas_label) for nda in ndas]
485-
nda_foregrounds = [nda if nda.numel() > 0 else torch.Tensor([0]) for nda in nda_foregrounds]
499+
nda_foregrounds = [nda if nda.numel() > 0 else MetaTensor([0.0]) for nda in nda_foregrounds]
486500

487501
unique_label = unique(ndas_label)
488502
if isinstance(ndas_label, (MetaTensor, torch.Tensor)):

monai/engines/utils.py

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -219,8 +219,8 @@ def __call__(
219219
`kwargs` supports other args for `Tensor.to()` API.
220220
"""
221221
image, label = default_prepare_batch(batchdata, device, non_blocking, **kwargs)
222-
args_ = list()
223-
kwargs_ = dict()
222+
args_ = []
223+
kwargs_ = {}
224224

225225
def _get_data(key: str) -> torch.Tensor:
226226
data = batchdata[key]

monai/losses/__init__.py

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -36,6 +36,7 @@
3636
from .giou_loss import BoxGIoULoss, giou
3737
from .hausdorff_loss import HausdorffDTLoss, LogHausdorffDTLoss
3838
from .image_dissimilarity import GlobalMutualInformationLoss, LocalNormalizedCrossCorrelationLoss
39+
from .mcc_loss import MCCLoss
3940
from .multi_scale import MultiScaleLoss
4041
from .nacl_loss import NACLLoss
4142
from .perceptual import PerceptualLoss

monai/losses/mcc_loss.py

Lines changed: 188 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,188 @@
1+
# Copyright (c) MONAI Consortium
2+
# Licensed under the Apache License, Version 2.0 (the "License");
3+
# you may not use this file except in compliance with the License.
4+
# You may obtain a copy of the License at
5+
# http://www.apache.org/licenses/LICENSE-2.0
6+
# Unless required by applicable law or agreed to in writing, software
7+
# distributed under the License is distributed on an "AS IS" BASIS,
8+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
9+
# See the License for the specific language governing permissions and
10+
# limitations under the License.
11+
12+
from __future__ import annotations
13+
14+
import warnings
15+
from collections.abc import Callable
16+
17+
import torch
18+
from torch.nn.modules.loss import _Loss
19+
20+
from monai.networks import one_hot
21+
from monai.utils import LossReduction
22+
23+
24+
class MCCLoss(_Loss):
25+
"""
26+
Compute the Matthews Correlation Coefficient (MCC) loss between two tensors.
27+
28+
Unlike Dice and Tversky losses which only use TP, FP, and FN, the MCC loss considers all four entries
29+
of the confusion matrix (TP, TN, FP, FN), making it effective for class-imbalanced segmentation tasks
30+
where background dominates the image. The loss is computed as ``1 - MCC`` where
31+
``MCC = (TP * TN - FP * FN) / sqrt((TP+FP)(TP+FN)(TN+FP)(TN+FN))``.
32+
33+
The soft confusion matrix entries are computed as:
34+
35+
- ``TP = sum(input * target)``
36+
- ``TN = sum((1 - input) * (1 - target))``
37+
- ``FP = sum(input * (1 - target))``
38+
- ``FN = sum((1 - input) * target)``
39+
40+
The data `input` (BNHW[D] where N is number of classes) is compared with ground truth `target` (BNHW[D]).
41+
42+
Note that axis N of `input` is expected to be logits or probabilities for each class, if passing logits as input,
43+
must set `sigmoid=True` or `softmax=True`, or specifying `other_act`. And the same axis of `target`
44+
can be 1 or N (one-hot format).
45+
46+
The original paper:
47+
48+
Abhishek, K. and Hamarneh, G. (2021) Matthews Correlation Coefficient Loss for Deep Convolutional
49+
Networks: Application to Skin Lesion Segmentation. IEEE ISBI, pp. 225-229.
50+
(https://doi.org/10.1109/ISBI48211.2021.9433782)
51+
52+
"""
53+
54+
def __init__(
55+
self,
56+
include_background: bool = True,
57+
to_onehot_y: bool = False,
58+
sigmoid: bool = False,
59+
softmax: bool = False,
60+
other_act: Callable | None = None,
61+
reduction: LossReduction | str = LossReduction.MEAN,
62+
smooth_nr: float = 0.0,
63+
smooth_dr: float = 1e-5,
64+
batch: bool = False,
65+
) -> None:
66+
"""
67+
Args:
68+
include_background: if False, channel index 0 (background category) is excluded from the calculation.
69+
if the non-background segmentations are small compared to the total image size they can get
70+
overwhelmed by the signal from the background so excluding it in such cases helps convergence.
71+
to_onehot_y: whether to convert the ``target`` into the one-hot format,
72+
using the number of classes inferred from `input` (``input.shape[1]``). Defaults to False.
73+
sigmoid: if True, apply a sigmoid function to the prediction.
74+
softmax: if True, apply a softmax function to the prediction.
75+
other_act: callable function to execute other activation layers, Defaults to ``None``. for example:
76+
``other_act = torch.tanh``.
77+
reduction: {``"none"``, ``"mean"``, ``"sum"``}
78+
Specifies the reduction to apply to the output. Defaults to ``"mean"``.
79+
80+
- ``"none"``: no reduction will be applied.
81+
- ``"mean"``: the sum of the output will be divided by the number of elements in the output.
82+
- ``"sum"``: the output will be summed.
83+
84+
smooth_nr: a small constant added to the numerator to avoid zero.
85+
smooth_dr: a small constant added to the denominator to avoid nan.
86+
batch: whether to sum the confusion matrix entries over the batch dimension before computing MCC.
87+
Defaults to False, MCC is computed independently for each item in the batch
88+
before any `reduction`.
89+
90+
Raises:
91+
TypeError: When ``other_act`` is not an ``Optional[Callable]``.
92+
ValueError: When more than 1 of [``sigmoid=True``, ``softmax=True``, ``other_act is not None``].
93+
Incompatible values.
94+
95+
"""
96+
super().__init__(reduction=LossReduction(reduction).value)
97+
if other_act is not None and not callable(other_act):
98+
raise TypeError(f"other_act must be None or callable but is {type(other_act).__name__}.")
99+
if int(sigmoid) + int(softmax) + int(other_act is not None) > 1:
100+
raise ValueError("Incompatible values: more than 1 of [sigmoid=True, softmax=True, other_act is not None].")
101+
self.include_background = include_background
102+
self.to_onehot_y = to_onehot_y
103+
self.sigmoid = sigmoid
104+
self.softmax = softmax
105+
self.other_act = other_act
106+
self.smooth_nr = float(smooth_nr)
107+
self.smooth_dr = float(smooth_dr)
108+
self.batch = batch
109+
110+
def forward(self, input: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
111+
"""
112+
Args:
113+
input: the shape should be BNH[WD], where N is the number of classes.
114+
target: the shape should be BNH[WD] or B1H[WD], where N is the number of classes.
115+
116+
Raises:
117+
AssertionError: When input and target (after one hot transform if set)
118+
have different shapes.
119+
ValueError: When ``self.reduction`` is not one of ["mean", "sum", "none"].
120+
121+
Example:
122+
>>> from monai.losses.mcc_loss import MCCLoss
123+
>>> import torch
124+
>>> B, C, H, W = 7, 1, 3, 2
125+
>>> input = torch.rand(B, C, H, W)
126+
>>> target = torch.randint(low=0, high=2, size=(B, C, H, W)).float()
127+
>>> self = MCCLoss(reduction='none')
128+
>>> loss = self(input, target)
129+
"""
130+
if self.sigmoid:
131+
input = torch.sigmoid(input)
132+
133+
n_pred_ch = input.shape[1]
134+
if self.softmax:
135+
if n_pred_ch == 1:
136+
warnings.warn("single channel prediction, `softmax=True` ignored.")
137+
else:
138+
input = torch.softmax(input, 1)
139+
140+
if self.other_act is not None:
141+
input = self.other_act(input)
142+
143+
if self.to_onehot_y:
144+
if n_pred_ch == 1:
145+
warnings.warn("single channel prediction, `to_onehot_y=True` ignored.")
146+
else:
147+
target = one_hot(target, num_classes=n_pred_ch)
148+
149+
if not self.include_background:
150+
if n_pred_ch == 1:
151+
warnings.warn("single channel prediction, `include_background=False` ignored.")
152+
else:
153+
target = target[:, 1:]
154+
input = input[:, 1:]
155+
156+
if target.shape != input.shape:
157+
raise AssertionError(f"ground truth has differing shape ({target.shape}) from input ({input.shape})")
158+
159+
# reducing only spatial dimensions (not batch nor channels)
160+
reduce_axis: list[int] = torch.arange(2, len(input.shape)).tolist()
161+
if self.batch:
162+
reduce_axis = [0] + reduce_axis
163+
164+
# Soft confusion matrix entries (Eq. 5 in the paper).
165+
tp = torch.sum(input * target, dim=reduce_axis)
166+
tn = torch.sum((1.0 - input) * (1.0 - target), dim=reduce_axis)
167+
fp = torch.sum(input * (1.0 - target), dim=reduce_axis)
168+
fn = torch.sum((1.0 - input) * target, dim=reduce_axis)
169+
170+
# MCC (Eq. 3) and loss (Eq. 4).
171+
numerator = tp * tn - fp * fn + self.smooth_nr
172+
denominator = torch.sqrt((tp + fp) * (tp + fn) * (tn + fp) * (tn + fn) + self.smooth_dr)
173+
174+
mcc = numerator / denominator
175+
score: torch.Tensor = 1.0 - mcc
176+
177+
# When fp = fn = 0, prediction is perfect but the denominator product
178+
# tends to 0 when tp = 0 or tn = 0, giving mcc ~ 0 instead of 1.
179+
perfect = (fp == 0) & (fn == 0)
180+
score = torch.where(perfect, torch.zeros_like(score), score)
181+
182+
if self.reduction == LossReduction.SUM.value:
183+
return torch.sum(score)
184+
if self.reduction == LossReduction.NONE.value:
185+
return score
186+
if self.reduction == LossReduction.MEAN.value:
187+
return torch.mean(score)
188+
raise ValueError(f'Unsupported reduction: {self.reduction}, available options are ["mean", "sum", "none"].')

monai/transforms/inverse.py

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -282,8 +282,8 @@ def track_transform_meta(
282282
msg += f" for key {key}"
283283

284284
pend = out_obj.pending_operations[-1]
285-
statuses = pend.get(TraceKeys.STATUSES, dict())
286-
messages = statuses.get(TraceStatusKeys.PENDING_DURING_APPLY, list())
285+
statuses = pend.get(TraceKeys.STATUSES, {})
286+
messages = statuses.get(TraceStatusKeys.PENDING_DURING_APPLY, [])
287287
messages.append(msg)
288288
statuses[TraceStatusKeys.PENDING_DURING_APPLY] = messages
289289
info[TraceKeys.STATUSES] = statuses

requirements-dev.txt

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -14,9 +14,9 @@ mccabe
1414
pep8-naming
1515
pycodestyle
1616
pyflakes
17-
black>=25.1.0
18-
isort>=5.1, !=6.0.0
19-
ruff
17+
black==25.1.0
18+
isort>=5.1, <6, !=6.0.0
19+
ruff>=0.14.11,<0.15
2020
pytype>=2020.6.1, <=2024.4.11; platform_system != "Windows"
2121
types-setuptools
2222
mypy>=1.5.0, <1.12.0

0 commit comments

Comments
 (0)