DICOM Images

LightlyTrain supports loading DICOM images for pretraining, distillation, and fine-tuning.

Note

Currently, we do not support loading DICOM images as segmentation masks.

PyDICOM Support

Under the hood, LightlyTrain uses the pydicom library to read and process DICOM images. It is included as an optional dependency.

To install LightlyTrain with PyDICOM support, do:

pip install lightly-train[dicom]

For DICOM images that may require additional processing, LightlyTrain automatically applies the following using pydicom functions:

• converting color space from YBR to RGB via convert_color_space

• decoding palette color images to RGB via apply_color_lut

• rescaling images to HU values via apply_modality_lut

Please refer to the respective pydicom documentation for more details on these functions.

Supported Image Types

The following DICOM image types listed in pydicom.examples are supported:

Type	SOP Class	num_channels
ct	CT Image	1
mr	MR Image	1
overlay	MR Image	1
rgb_color	US Image	3
palette_color	US Image	1
jpeg2k	US Image	3

Currently, LightlyTrain loads one DICOM file as one image. Combining slices from multiple DICOM files into a 3D volume is not supported. As a result, RT Dose (rt_dose), ECG Waveform (waveform), and US Multi-frame Image (ybr_color) are not supported.

Transforms

When training with DICOM images, you may need to customize the applied transforms for medical domains. We recommend keeping spatial operations such as RandomResize, RandomFlip, and RandomRotation. We strongly suggest disabling the following transforms—even for 3D RGB DICOM files—as they often do not make sense for medical images:

• Solarize

• RandomGrayScale

• ColorJitter

• ChannelDrop

Disable them by setting the corresponding transform argument to None:

transform_args={
    "solarize": None,
    "random_grayscale": None,
    "color_jitter": None,
    "channel_drop": None,
},

Also be aware that some transform options may be inappropriate depending on the acquisition protocol. For example, horizontal flips in RandomFlip may not be suitable for certain medical images. To disable horizontal flips:

transform_args={
    "random_flip": {
        "horizontal_prob": 0.0,
    },
},