Using DICOM Files as Input for a 3D CNN

I am working on a project with tomosynthesis images of patients with breast cancer, in which I will use 3D CNNs. At the moment, I am selecting the slices where the cancer is visible, and my file format is DICOM. To use 3D CNNs, should I save the selected slices in a new single DICOM file, or should I save each slice in a separate file? What is the best format to use as input for my 3D CNN?

Additionally, to save the selected slices, which DICOM file metadata should I modify? I am modifying only the SOPInstanceUID and the PixelData.

Answer 1

[aditi bagora]

Hi @Raquel,

I would suggest to create a separate DICOM file with the selected images and then pass this 3D pixel data i.e. stack of images of selected slices as an input to the 3D CNN. In order to create a new file along with the pixeldata following metdata fields needs to be updated:

• SeriesInstanceUID : Identifier for a new series
• SOPInstanceUID: Within the series identifies the instance
• InstanceNumber: Identifies the slice number in stack of slices.

However, if your goal is for the 3D CNN to learn how to classify slices based on the presence of cancer, it might be beneficial not to filter the data. Providing the entire series, including slices both with and without visible cancer, will enable the CNN to learn to distinguish between slices that contain cancer and those that do not. Otherwise, by only supplying positive samples, you limit the model's ability to differentiate.