How do I use BoundingRectd?

[sarbabi]

I want to crop the image based on the bounding box of roi.
I do this:

BoundingRectd(keys=["roi"]),
Cropd(keys=["image", "roi", "label"]
      , roi_start=[roi_bbox[0][0], roi_bbox[0][1], roi_bbox[0][2]]
      , roi_end=[roi_bbox[1][0], roi_bbox[1][1], roi_bbox[1][2]]
     ),

according to documentation "bbox_key_postfix (str) – the output bounding box coordinates will be written to the value of {key}_{bbox_key_postfix}", so it should be roi_bbox? or how can I access the bounding box in next transform?

Thanks in advance,

[KumoLiu]

Hi @sarbabi, if you want to crop the image based on the bounding box of roi, you can simply use CropForeground like this:

CropForeground(keys=["images"], source_key="roi")

[KumoLiu]

Hi, this is a simple example of how to use BoundingRectd. Hope it can help you solve your problem.

from monai.transforms import BoundingRectd
from monai.data.synthetic import create_test_image_3d

image, label = create_test_image_3d(height=100, width=100, depth=100, num_objs=1, rad_max=30, rad_min=5, channel_dim=0)
print("Generated data shape:", label.shape)

item = BoundingRectd(keys="label", bbox_key_postfix="bbox")({"label": label})
print("Boundingbox:", item["label_bbox"])

[sarbabi]

I was wondering is it possible to use it sequentially with other transforms, like the way they are passed in Compose:

train_transforms = Compose(
    [
        LoadImaged(keys=["image", "roi", "label"]),
        EnsureChannelFirstd(keys=["image", "roi", "label"]),
        Orientationd(keys=["image", "roi", "label"], axcodes="RAS"),
        CropForegroundd(keys=["image", "roi", "label"], source_key="image"),
        SpatialCropd(keys=["image", "roi", "label"], roi_start=[0,0,0], roi_end=[480,135, 400]),
        ScaleIntensityRangePercentilesd(keys=["image"], lower=0, upper=100, b_min=-1, b_max=1),
        item = BoundingRectd(keys=["roi"], bbox_key_postfix="bbox"),
        Cropd(keys=["image", "roi", "label"]
              , roi_start=[item["roi_bbox"][0][0], item["roi_bbox"][0][1], item["roi_bbox"][0][2]]
              , roi_end=[item["roi_bbox"][1][0], item["roi_bbox"][1][1], item["roi_bbox"][1][2]]
             ),
        
    ]
)

This one brings up error:

    item = BoundingRectd(keys=["roi"], bbox_key_postfix="bbox"),
    ^
SyntaxError: invalid syntax. Maybe you meant '==' or ':=' instead of '='?

[KumoLiu]

Hi, from your given example, I think CropForeground can actually solve your problem. BoundingRect is not suitable in this case.

[sarbabi]

The idea is to crop the same size from all images that always contains roi, something like this is what I'm looking for. but it doesn't work this way. Any solution?

        item = BoundingRectd(keys=["roi"], bbox_key_postfix="bbox"),
        Cropd(keys=["image", "roi", "label"]
              , roi_start=[item["roi_bbox"][0][0], item["roi_bbox"][0][1], item["roi_bbox"][0][2]]
              , roi_end=[item["roi_bbox"][0][0]+80, item["roi_bbox"][0][1]+32, item["roi_bbox"][0][2]+96]
             ),

[wyli]

I think @KumoLiu's idea is

train_transforms = Compose(
    [
        LoadImaged(keys=["image", "roi", "label"]),
        EnsureChannelFirstd(keys=["image", "roi", "label"]),
        Orientationd(keys=["image", "roi", "label"], axcodes="RAS"),
        CropForegroundd(keys=["image", "roi", "label"], source_key="image"),
        SpatialCropd(keys=["image", "roi", "label"], roi_start=[0,0,0], roi_end=[480,135, 400]),
        ScaleIntensityRangePercentilesd(keys=["image"], lower=0, upper=100, b_min=-1, b_max=1),
        CropForegroundd(keys=["image", "roi", "label"], source_key="roi"),
    ]
)

or you'll need a customized transform (not tested, just to show the idea):

class MyCrop(monai.transforms.MapTransform):
    def __call__(self, data_dict):
        data_dict["image"] = SpatialCrop(roi_start=data_dict["roi_bbox"]..., roi_end=data_dict["roi_bbox"]...)(data_dict["image"])
        data_dict["label"] = SpatialCrop(roi_start=data_dict["roi_bbox"]..., roi_end=data_dict["roi_bbox"]...)(data_dict["label"])
        return data_dict

train_transforms = Compose(
    [
        LoadImaged(keys=["image", "roi", "label"]),
        EnsureChannelFirstd(keys=["image", "roi", "label"]),
        Orientationd(keys=["image", "roi", "label"], axcodes="RAS"),
        BoundingRectd(keys=["roi"], bbox_key_postfix="bbox"),
        MyCrop(keys=["image", "label"]),
    ]
)