Keywords
automated image analysis - segmentation - CT scan brain - traumatic brain injury
Introduction
Traumatic brain injury (TBI) refers to the neuropathological changes and brain dysfunction
due to any injury to the head. TBI has emerged as a silent epidemic predominantly
affecting the young and productive population and adds to the mortality, morbidity,
and societal burden.[1] CT head plain is the investigation of choice in patients with a head injury. The
complexity and dynamic nature of TBI necessitate prompt and accurate identification
of pathology on CT images and subsequent appropriate management for optimal outcome.
Therefore, it is crucial to have facilities for CT scans, and at the same time, it
is essential to have personnel trained to accurately interpret the CT image findings.
Imaging is an important clinical tool used in the management of patients with brain
injury. The objectives of the present study are to propose an algorithm for automated
image segmentation and interpretation of CT scans of the brain and propose an algorithm
for identification and categorization of the abnormal CT findings in patients with
TBI.
Characteristics of Brain Pathology to be Identified
Characteristics of Brain Pathology to be Identified
Important pathological findings on the CT brain are tabulated in [Table 1].[2] Each can be categorized based on the CT as shown earlier based on the zones described
have characteristic appearance as illustrated below. For example, on CT imaging, the
acute EDH appears as a well-defined, hyperdense, biconvex, extra-axial collection.
It is usually associated with an overlying skull fracture. Mass effect with sulcal
effacement and midline shift is frequently seen. Because the EDH is located in the
potential space between the dura and inner table of the skull, it rarely crosses cranial
sutures because the periosteal layer of the dura is firmly attached at sutural margins.
However, at the vertex, where the periosteum that forms the outer wall of the sagittal
sinus is not tightly attached to the sagittal suture, the EDH can cross midline. An
important imaging finding that predicts rapid expansion of an arterial EDH is the
presence of low-attenuation areas within the hyperdense hematoma (the so-called “swirl
sign”), thought to represent active bleeding. It is an ominous sign that needs to
be followed closely.[3]
[4]
Table 1
Based on lesion characteristics,[2] proposed zones on CT imaging
|
Zone
|
Definitions
|
Probable lesions
|
|
Zone I
|
Extracranial including scalp
|
Scalp lacerations
Swelling
|
|
Zone II
|
Skull
|
Various fractures
|
|
Zone III
|
Just inside the skull
Including dura mater, subdural space, subarachnoid space and adjacent brain parenchyma
|
Extradural hemorrhage (EDH)
Subdural hemorrhage (SDH, acute and chronic)
Subarachnoid hemorrhage (SAH)
Surface contusions
|
|
Zone IV
|
Cerebral gray and white matter, outside the central ventricular region
|
(Intracerebral hemorrhage (ICH)
Cerebral contusion
Cerebral, edema
|
|
Zone V
|
Central ventricular regions and adjacent brain parenchyma
|
Intraventricular hemorrhage (IVH)
Contusions
|
The Proposed Method for Automated CT Interpretation
The Proposed Method for Automated CT Interpretation
We propose image segmentation algorithms that can be applied for the precise detection
of TBI pathology ([Table 1]). Such detection can be helpful in further quantitative analysis of critical characteristics,
such as the size or volume of the lesion. Our proposed method is divided into three
main stages: image preprocessing, brain segmentation, and hydrocephalus segmentation.
Image Preprocessing
Each image from the input dataset will be normalized to a common intensity range in
this step. For brain segmentation, pixel intensities will be transformed to Hounsfield
units[1] (range from -1024 to ∼3071). According to the DICOM specification, every pixel in
the image will be scaled. After scaling image intensities, using other DICOM header
information, such as Window width and Window center, pixel intensities will be transformed from signed to unsigned values without quality
changes. In the case of CT images, 12 bits are sufficient to cover a whole range of
intensities. The data are shifted when using unsigned shorts, so all CT intensities
become positive numbers ranging from 0 to 4095 (gmin = 0, gmax = 4095).
Brain Segmentation
The second step will be aimed at the extraction of the whole brain. This step is necessary
for further quantitative assessment of the disease progress. Pixel transformation
performed in the previous step significantly increased the image's contrast. As a
result, the skull area and the CT scanner tube elements could be easily removed by
suppressing (setting to zero) any pixel in the image above 95% of the maximum pixel
intensity value. The selected threshold will be chosen empirically based on observing
the distribution of pixel intensities after their transformation. After removing the
skull and CT tube, extraction of the whole brain area was possible. For this purpose,
the 2D segmentation algorithm based on region growing was applied. This method requires
the selection of the initial seed point. It was decided to locate the seed at the
center of each cross-section as it is always contained in the brain area. The desired
region originates from the exact location of this point. Then, the region grows from
the seed point to adjacent points depending on the selected threshold. Threshold value
determines the scope of permissible difference of intensity between intensity of the
candidate pixel and an average intensity of pixels already classified into the region.
In the present method, we propose five zones on the CT scan from outside to inside
and different pathologies described on each zone. These zones are described in [Table 1] along with the corresponding lesions. The normal unsegmented CT brain plain image
is shown in [Fig. 1A]. [Fig. 1B] and [1C] shows the segmented zone 1 corresponding to the scalp and skull. The proposed five
zones are shown in [Fig. 2].
Fig. 1
(A): Normal unsegmented plain CT brain. (B) Segmented zone of scalp and skull in sagittal view. (C) Coronal View.
Fig. 2 Proposed five zones from outside to inside comprising scalp as zone 1, skull as zone
II, subdural space as zone III, cerebral gray matter as zone IV, central ventricular
regions and adjacent parenchyma as zone V.
Advantages and Challenges
Advantages and Challenges
The advantage of developing this algorithm is prompt identification of pathology on
CT scan by a primary physician and nonneurosurgeon. Early and accurate CT scan interpretation
will facilitate the timely transfer of TBI patients to a center equipped with facilities
for TBI care and allow for the appropriate treatment initiation at the first point
of contact. The main disadvantage of this approach is that seeded region-growing algorithms
may not clearly define the boundaries of the stroke region. Moreover, to date, only
a single study has addressed the problem of detecting both hemorrhagic and ischemic
strokes in a given CT volume.[5] It is worth mentioning, however, that the availability of a high-quality template
does not in itself ensure a successful spatial normalization.[6]
Conclusion
Rapid advancements in medical imaging technology have resulted in accurate and early
diagnosis of many diseases conditions, better management planning, and improved outcomes
in neurosurgical practice.[7] Several authors have developed algorithms for automated analysis and segmentation
of CT images to interpret the findings as an adjunct to manual image reading.[7]
[8]
[9]
[10]
[11]
[12] The present article presents a conceptual analysis to explore the feasibility of
automated image analysis and segmentation to interpret the CT images in patients with
TBI.
