diff --git a/nibabel/nicom/csareader.py b/nibabel/nicom/csareader.py
index df379e0be..dd081b22c 100644
--- a/nibabel/nicom/csareader.py
+++ b/nibabel/nicom/csareader.py
@@ -1,6 +1,7 @@
 """CSA header reader from SPM spec"""
 
 import numpy as np
+import struct
 
 from .structreader import Unpacker
 from .utils import find_private_section
@@ -28,6 +29,10 @@ class CSAReadError(CSAError):
     pass
 
 
+class CSAWriteError(CSAError):
+    pass
+
+
 def get_csa_header(dcm_data, csa_type='image'):
     """Get CSA header information from DICOM header
 
@@ -161,6 +166,96 @@ def read(csa_str):
     return csa_dict
 
 
+def write(csa_header):
+    ''' Write string from CSA header `csa_header`
+
+    Parameters
+    ----------
+    csa_header : dict
+        header information as dict, where `header` has fields (at least)
+        ``type, n_tags, tags``.  ``header['tags']`` is also a dictionary
+        with one key, value pair for each tag in the header.
+
+    Returns
+    -------
+    csa_str : str
+        byte string containing CSA header information
+    '''
+    result = []
+    if csa_header['type'] == 2:
+        result.append(b'SV10')
+        result.append(csa_header['unused0'])
+    if not 0 < csa_header['n_tags'] <= 128:
+        raise CSAWriteError('Number of tags `t` should be '
+                            '0 < t <= 128')
+    result.append(struct.pack('2I',
+                              csa_header['n_tags'],
+                              csa_header['check'])
+                  )
+
+    # Build list of tags in correct order
+    tags = list(csa_header['tags'].items())
+    tags.sort(key=lambda x: x[1]['tag_no'])
+    tag0_n_items = tags[0][1]['n_items']
+
+    # Add the information for each tag
+    for tag_name, tag_dict in tags:
+        vm = tag_dict['vm']
+        vr = tag_dict['vr']
+        n_items = tag_dict['n_items']
+        assert n_items < 100
+        result.append(struct.pack('64si4s3i',
+                                  make_nt_str(tag_name),
+                                  vm,
+                                  make_nt_str(vr),
+                                  tag_dict['syngodt'],
+                                  n_items,
+                                  tag_dict['last3'])
+                      )
+
+        # Figure out the number of values for this tag
+        if vm == 0:
+            n_values = n_items
+        else:
+            n_values = vm
+
+        # Add each item for this tag
+        for item_no in range(n_items):
+            # Figure out the item length
+            if item_no >= n_values or tag_dict['items'][item_no] == '':
+                item_len = 0
+            else:
+                item = tag_dict['items'][item_no]
+                if not isinstance(item, str):
+                    item = str(item)
+                item_nt_str = make_nt_str(item)
+                item_len = len(item_nt_str)
+
+            # These values aren't actually preserved in the dict
+            # representation of the header. Best we can do is set the ones
+            # that determine the item length appropriately.
+            x0, x1, x2, x3 = 0, 0, 0, 0
+            if csa_header['type'] == 1:  # CSA1 - odd length calculation
+                x0 = tag0_n_items + item_len
+                if item_len < 0 or (ptr + item_len) > csa_len:
+                    if item_no < vm:
+                        items.append('')
+                    break
+            else:  # CSA2
+                x1 = item_len
+            result.append(struct.pack('4i', x0, x1, x2, x3))
+
+            if item_len == 0:
+                continue
+
+            result.append(item_nt_str)
+            # go to 4 byte boundary
+            plus4 = item_len % 4
+            if plus4 != 0:
+                result.append(b'\x00' * (4 - plus4))
+    return b''.join(result)
+
+
 def get_scalar(csa_dict, tag_name):
     try:
         items = csa_dict['tags'][tag_name]['items']
@@ -258,3 +353,18 @@ def nt_str(s):
     if zero_pos == -1:
         return s
     return s[:zero_pos].decode('latin-1')
+
+
+def make_nt_str(s):
+    ''' Create a null terminated byte string from a unicode object.
+
+    Parameters
+    ----------
+    s : unicode
+
+    Returns
+    -------
+    result : bytes
+       s encoded as latin-1 with a null char appended
+    '''
+    return s.encode('latin-1') + b'\x00'
diff --git a/nibabel/nicom/tests/test_csareader.py b/nibabel/nicom/tests/test_csareader.py
index f31f4a393..67ae44ecb 100644
--- a/nibabel/nicom/tests/test_csareader.py
+++ b/nibabel/nicom/tests/test_csareader.py
@@ -130,3 +130,14 @@ def test_missing_csa_elem():
     del dcm[csa_tag]
     hdr = csa.get_csa_header(dcm, 'image')
     assert hdr is None
+
+
+def test_read_write_rt():
+    # Try doing a read-write-read round trip and make sure the dictionary
+    # representation of the header is the same. We can't exactly reproduce the
+    # original string representation currently.
+    for csa_str in (CSA2_B0, CSA2_B1000):
+        csa_info = csa.read(csa_str)
+        new_csa_str = csa.write(csa_info)
+        new_csa_info = csa.read(new_csa_str)
+        assert csa_info == new_csa_info