diff --git a/README.md b/README.md
index 63e4518..79f5b6b 100644
--- a/README.md
+++ b/README.md
@@ -42,7 +42,7 @@ print("First segment info:\n" + json.dumps(segment0, sort_keys=False, indent=4))
 
 #### Extract segments by terminology
 
-Example for getting a 3D NRRD file that has label values assigned based on [standard terminology codes](UsingStandardTerminology.md).
+Example for getting a 3D [NRRD](SegmentationFileFormat.md) file that has label values assigned based on [standard terminology codes](UsingStandardTerminology.md).
 Terminology is a `dict` that must specify `category` and `type` codes and may optionally also specify `typeModifier`, `anatomicRegion`, and `anatomicRegionModifier`. Each code is specifed by a triplet of "coding scheme designator", "code value", "code meaning" in a list.
 
 Coding scheme designator is typically `SCT` (SNOMED-CT) for clinical images. You can find codes in the [SNOMED-CT browser](https://browser.ihtsdotools.org/). When code exists for "entire X" and "structure of X" then always use the "structure" code ("entire" code has a very strict meaning that is rarely applicable in practice).
diff --git a/SegmentationFileFormat.md b/SegmentationFileFormat.md
new file mode 100644
index 0000000..1089705
--- /dev/null
+++ b/SegmentationFileFormat.md
@@ -0,0 +1,51 @@
+# File formats for image segmentations
+
+## Commonly used file formats
+
+### DICOM
+
+DICOM allows saving image segmentations along with rich, structured metadata in a standard, very well documented way in [DICOM Segmentation](https://dicom.nema.org/dicom/2013/output/chtml/part03/sect_A.51.html).
+
+Pro:
+- Compatibility with clinical software: DICOM is universally supported by all clinical imaging softare, DICOM Segmentation is not very widely supported but its adoption is increasing.
+- Long-term stability: DICOM format is changed very slowly and carefully
+- Rich, structured metadata storage, all defined carefully in the DICOM standard
+
+Con:
+- Standard is very large, complex
+- Low performance: size of segmentation metadata may be huge (magnitudes larger than the segmentation data itself), encoding and decoding can be magnitudes slower compared to simple 3D image formats
+
+### NRRD
+
+Very simple and efficient file format for storing multi-dimensional images. 3D Slicer and related software established convention for [storing segmentation metadata](https://slicer.readthedocs.io/en/latest/developer_guide/modules/segmentations.html#segmentation-labelmap-file-format-seg-nrrd) (commonly referred to as `.seg.nrrd` files).
+
+Pro:
+- Very simple
+- Compatibility with research software and toolkits: most libraries and research application support this file format
+- Common convention exists for storing essential metadata
+- File header is human-readable: easily readable, modifiable with any text editor
+
+Con:
+- Not compatible with clinical software
+- Common conventions for metadata storage only include essential metadata (not as rich or structured as DICOM)
+
+### NIFTI
+
+Very limited, yet complex and ambiguous file format. Unfortunately, for historical reasons, it is commonly used in medical image computing.
+
+Pro:
+- Compatibility with research software and toolkits: most libraries and research application support this file format
+- Somewhat simple: simpler than DICOM, but much more complex than NRRD
+- It can store some neuroimaging-specific metadata in a standard way: this makes the file format beneficial for neuroimaging
+
+Con:
+- Orientation definition in NIFTI files can be ambiguous: there are multiple ways to define orientation, they can be both present, and contain contradicting information. Various softare will interpret these ambiguous files differently. See for example these discussions: [1](https://www.openfmri.org/dataset-orientation-issues/), [2](https://discourse.slicer.org/t/orientation-origin/23365), [3](https://github.com/spinalcordtoolbox/spinalcordtoolbox/issues/3283)
+- Not compatible with clinical software
+- Common convention for storing essential metadata does not exist
+- File header is not human-readable: you cannot use a text editor/viewer to see the actual file header but you always have to rely on a parser to give you an interpretation of the file header
+
+## Recommended file formats
+
+- DICOM format is recommended for long-term archival due to compatibility and stability of the file format and the ability to encode rich, structured metadata in a well-defined manner.
+- NRRD file format is recommended for local storage and data exchange, due to its simplicity, efficiency, and non-ambiguity.
+- NIFTI format usage may be justified for neuroimaging, as many neuroimaging pipelines only support this format.