Supporting DICOM via pydicom

digits/dataset/images/classification/job.py

@@ -149,6 +149,7 @@ def json_dict(self, verbose=False):
                     "image_width":      t.image_dims[0],
                     "image_height":     t.image_dims[1],
                     "image_channels":   t.image_dims[2],
+                    "image_bpp":        t.resize_bpp,
                     "backend":          t.backend,
                     "encoding":         t.encoding,
                     "compression":      t.compression,

digits/dataset/images/classification/views.py

@@ -109,6 +109,7 @@ def from_folders(job, form):
                 backend     = backend,
                 image_dims  = job.image_dims,
                 resize_mode = job.resize_mode,
+                resize_bpp = job.resize_bpp,
                 encoding    = encoding,
                 compression = compression,
                 mean_file   = utils.constants.MEAN_FILE_CAFFE,
@@ -126,6 +127,7 @@ def from_folders(job, form):
                     backend     = backend,
                     image_dims  = job.image_dims,
                     resize_mode = job.resize_mode,
+                    resize_bpp = job.resize_bpp,
                     encoding    = encoding,
                     compression = compression,
                     labels_file = job.labels_file,
@@ -142,6 +144,7 @@ def from_folders(job, form):
                     backend     = backend,
                     image_dims  = job.image_dims,
                     resize_mode = job.resize_mode,
+                    resize_bpp = job.resize_bpp,
                     encoding    = encoding,
                     compression = compression,
                     labels_file = job.labels_file,
@@ -188,6 +191,7 @@ def from_files(job, form):
                 image_dims  = job.image_dims,
                 image_folder= image_folder,
                 resize_mode = job.resize_mode,
+                resize_bpp = job.resize.bpp,
                 encoding    = encoding,
                 compression = compression,
                 mean_file   = utils.constants.MEAN_FILE_CAFFE,
@@ -220,6 +224,7 @@ def from_files(job, form):
                     image_dims  = job.image_dims,
                     image_folder= image_folder,
                     resize_mode = job.resize_mode,
+                    resize_bpp = job.resize_bpp,
                     encoding    = encoding,
                     compression = compression,
                     labels_file = job.labels_file,
@@ -251,6 +256,7 @@ def from_files(job, form):
                     image_dims  = job.image_dims,
                     image_folder= image_folder,
                     resize_mode = job.resize_mode,
+                    resize_bpp = job.resize_bpp,
                     encoding    = encoding,
                     compression = compression,
                     labels_file = job.labels_file,
@@ -303,7 +309,8 @@ def create():
                     int(form.resize_width.data),
                     int(form.resize_channels.data),
                     ),
-                resize_mode = form.resize_mode.data
+                resize_mode = form.resize_mode.data,
+                resize_bpp = int(form.resize_bpp.data)
                 )
 
         if form.method.data == 'folder':

digits/dataset/images/forms.py

@@ -44,3 +44,8 @@ class ImageDatasetForm(DatasetForm):
             choices=ImageDatasetJob.resize_mode_choices(),
             tooltip = "Options for dealing with aspect ratio changes during resize. See examples below."
             )
+    resize_bpp = utils.forms.SelectField(u'Bits per pixel',
+            default='8',
+            choices=[('8', '8-bit (color or grayscale)'), ('32', '32-bit floating point (grayscale only)')],
+            tooltip="Storing 32-bit floating point for certain medical images."
+            )

digits/dataset/images/job.py

@@ -19,6 +19,7 @@ def __init__(self, **kwargs):
         """
         self.image_dims = kwargs.pop('image_dims', None)
         self.resize_mode = kwargs.pop('resize_mode', None)
+        self.resize_bpp = kwargs.pop('resize_bpp', None)
 
         super(ImageDatasetJob, self).__init__(**kwargs)
         self.pickver_job_dataset_image = PICKLE_VERSION

digits/dataset/tasks/create_db.py

@@ -40,6 +40,7 @@ def __init__(self, input_file, db_name, backend, image_dims, **kwargs):
         self.shuffle = kwargs.pop('shuffle', True)
         self.resize_mode = kwargs.pop('resize_mode' , None)
         self.encoding = kwargs.pop('encoding', None)
+        self.resize_bpp = kwargs.pop('resize_bpp', None)
         self.compression = kwargs.pop('compression', None)
         self.mean_file = kwargs.pop('mean_file', None)
         self.labels_file = kwargs.pop('labels_file', None)
@@ -147,6 +148,7 @@ def task_arguments(self, resources, env):
                 '--backend=%s' % self.backend,
                 '--channels=%s' % self.image_dims[2],
                 '--resize_mode=%s' % self.resize_mode,
+                '--resize_bpp=%s' % self.resize_bpp
                 ]
 
         if self.mean_file is not None:

digits/templates/datasets/images/classification/new.html

@@ -56,6 +56,13 @@ <h1>New Image Classification Dataset</h1>
                     {{ form.resize_mode(class='form-control') }}
                 </div>
             </div>
+            <div class="row">
+                <div class="form-group{{mark_errors([form.resize_bpp])}}">
+                    {{ form.resize_bpp.label }}
+                    {{ form.resize_bpp.tooltip }}
+                    {{ form.resize_bpp(class='form-control') }}
+                </div>
+            </div>
             <div class="row">
                 <a class="btn btn-info" href="#" onClick="return showResizeExample();">See example</a>
                 <br>

digits/templates/datasets/images/classification/show.html

@@ -19,6 +19,8 @@ <h4>Job Information</h4>
         <dd>{{'Color' if job.image_dims[2] == 3 else 'Grayscale'}}</dd>
         <dt>Resize Transformation</dt>
         <dd>{{ job.resize_mode_name() }}</dd>
+        <dt>Image bit-depth</dt>
+        <dd>{{ job.resize_bpp }}</dd>
         <dt>DB Backend</dt>
         <dd>{{job.get_backend()}}</dd>
         <dt>Image Encoding</dt>

digits/templates/datasets/images/classification/summary.html

@@ -24,6 +24,8 @@ <h4>
     GRAYSCALE
     {% endif %}
     </dd>
+    <dt>Image bit-depth</dt>
+    <dd>{{dataset.resize_bpp}}</dd>
     <dt>DB backend</dt>
     <dd>{{dataset.get_backend()}}</dd>
     {% for task in dataset.create_db_tasks() %}

digits/templates/datasets/images/generic/summary.html

@@ -21,6 +21,7 @@ <h4>
         <li><b>Image Count</b> - {{task.image_count}}</li>
         <li><b>Image Dimensions</b> -
         {{task.image_width}}x{{task.image_height}}x{{task.image_channels}}</li>
+        <li><b>Image bit-depth</b></li>
     </ul>
     {% endfor %}
 </ul>

digits/tools/create_db.py

@@ -202,6 +202,7 @@ def create_db(input_file, output_dir,
         image_width, image_height, image_channels,
         backend,
         resize_mode     = None,
+        resize_bpp      = None,
         image_folder    = None,
         shuffle         = True,
         mean_files      = None,
@@ -241,6 +242,8 @@ def create_db(input_file, output_dir,
         raise ValueError('invalid number of channels')
     if resize_mode not in [None, 'crop', 'squash', 'fill', 'half_crop']:
         raise ValueError('invalid resize_mode')
+    if resize_bpp not in [None, '8', '32']:
+        raise ValueError('invalid resize_bpp')
     if image_folder is not None and not os.path.exists(image_folder):
         raise ValueError('image_folder does not exist')
     if mean_files:
@@ -269,11 +272,14 @@ def create_db(input_file, output_dir,
     write_queue = Queue.Queue(2*batch_size)
     summary_queue = Queue.Queue()
 
+    # Init helper function for notification between threads
+    _notification(reset=True)
+
     for _ in xrange(num_threads):
         p = threading.Thread(target=_load_thread,
                 args=(load_queue, write_queue, summary_queue,
                     image_width, image_height, image_channels,
-                    resize_mode, image_folder, compute_mean),
+                    resize_mode, image_folder, compute_mean, resize_bpp),
                 kwargs={'backend': backend,
                     'encoding': kwargs.get('encoding', None)},
                 )
@@ -331,6 +337,9 @@ def _create_lmdb(image_count, write_queue, batch_size, output_dir,
 
         processed_something = False
 
+        if _notification():
+            break
+
         if not summary_queue.empty():
             result_count, result_sum = summary_queue.get()
             images_loaded += result_count
@@ -360,6 +369,9 @@ def _create_lmdb(image_count, write_queue, batch_size, output_dir,
         _write_batch_lmdb(db, batch, images_written)
         images_written += len(batch)
 
+    if _notification():
+        raise WriteError('. '.join(_notification()))
+
     if images_loaded == 0:
         raise LoadError('no images loaded from input file')
     logger.debug('%s images loaded' % images_loaded)
@@ -498,6 +510,28 @@ def _fill_load_queue(filename, queue, shuffle):
 
     return valid_lines
 
+def _notification(reset=False, message=None):
+    """
+
+    Args:
+        reset: clear the message list if True
+        message: the error message
+
+    Returns:
+        False: if no message stored and not reset
+        The messages (a list): if some messages stored
+    """
+    if not reset:
+        if message is None:
+            if len(_notification.messages) == 0:
+                return False
+            else:
+                return _notification.messages
+        else:
+            _notification.messages.append(message)
+    else:
+        _notification.messages = list()
+
 def _parse_line(line, distribution):
     """
     Parse a line in the input file into (path, label)
@@ -541,7 +575,7 @@ def _calculate_num_threads(batch_size, shuffle):
 
 def _load_thread(load_queue, write_queue, summary_queue,
         image_width, image_height, image_channels,
-        resize_mode, image_folder, compute_mean,
+        resize_mode, image_folder, compute_mean, resize_bpp,
         backend=None, encoding=None):
     """
     Consumes items in load_queue
@@ -574,17 +608,21 @@ def _load_thread(load_queue, write_queue, summary_queue,
                 image_height, image_width,
                 channels    = image_channels,
                 resize_mode = resize_mode,
+                resize_bpp = resize_bpp
                 )
 
         if compute_mean:
             image_sum += image
 
-        if backend == 'lmdb':
-            datum = _array_to_datum(image, label, encoding)
-            write_queue.put(datum)
-        else:
-            write_queue.put((image, label))
-
+        try:
+            if backend == 'lmdb':
+                datum = _array_to_datum(image, label, encoding)
+                write_queue.put(datum)
+            else:
+                write_queue.put((image, label))
+        except IOError as e:  # report error to user (possibly save 16-bit image to PNG/JPG)
+            _notification(message=e.message)
+            break
         images_added += 1
 
     summary_queue.put((images_added, image_sum))
@@ -616,6 +654,8 @@ def _array_to_datum(image, label, encoding):
             image = image[np.newaxis,:,:]
         else:
             raise Exception('Image has unrecognized shape: "%s"' % image.shape)
+        if np.issubdtype(image.dtype, float):
+            image = image.astype(float)
         datum = caffe.io.array_to_datum(image, label)
     else:
         datum = caffe_pb2.Datum()
@@ -667,7 +707,7 @@ def _save_means(image_sum, image_count, mean_files):
     """
     Save mean[s] to file
     """
-    mean = np.around(image_sum / image_count).astype(np.uint8)
+    mean = np.around(image_sum / image_count).astype(np.float)
     for mean_file in mean_files:
         if mean_file.lower().endswith('.npy'):
             np.save(mean_file, mean)
@@ -693,7 +733,13 @@ def _save_means(image_sum, image_count, mean_files):
             with open(mean_file, 'wb') as outfile:
                 outfile.write(blob.SerializeToString())
         elif mean_file.lower().endswith(('.jpg', '.jpeg', '.png')):
-            image = PIL.Image.fromarray(mean)
+            #ensure pixel range is within supported format
+            if mean.max() < 256:  # works for three formats
+                image = PIL.Image.fromarray(mean.astype(np.uint8))
+            elif mean_file.lower().endswith('.png'):  # png supports higher color depth
+                image = PIL.Image.fromarray(mean).convert('I')
+            else:  # reduce color depth for jpg or jpeg
+                image = PIL.Image.fromarray(mean*255/mean.max()).convert('L')
             image.save(mean_file)
         else:
             logger.warning('Unrecognized file extension for mean file: "%s"' % mean_file)
@@ -730,6 +776,9 @@ def _save_means(image_sum, image_count, mean_files):
     parser.add_argument('-r', '--resize_mode',
             help='resize mode for images (must be "crop", "squash" [default], "fill" or "half_crop")'
             )
+    parser.add_argument('--resize_bpp',
+            help='bit per pixel for resized images (must be 8 (color/grayscale) or 32 (grayscale only)")'
+            )
     parser.add_argument('-m', '--mean_file', action='append',
             help="location to output the image mean (doesn't save mean if not specified)")
     parser.add_argument('-f', '--image_folder',
@@ -766,6 +815,7 @@ def _save_means(image_sum, image_count, mean_files):
                 args['width'], args['height'], args['channels'],
                 args['backend'],
                 resize_mode     = args['resize_mode'],
+                resize_bpp      = args['resize_bpp'],
                 image_folder    = args['image_folder'],
                 shuffle         = args['shuffle'],
                 mean_files      = args['mean_file'],