1.

Introduction

Traumatic brain injury (TBI) is a leading cause of morbidity and mortality with an estimated 1.6 million hospitalized and 66,000 deceased patients in Europe every year.¹ Computed tomography (CT) is generally performed as a primary standard imaging modality for acute TBI.² The advantages of CT include large-scale availability and short preparation and study time.^3,4 CT reliably detects intracranial hemorrhages and contusions, bone pathology, cerebral edema, and incipient herniation.^4,5 Limitations of noncontrast CT include a poor sensitivity for small hemorrhages, diffuse axonal injury (DAI), arterial dissection, and vascular damage.^6,7 In particular, in TBI patients with secondary progression of initially undetected small cerebral hemorrhages might cause a decline in the functional status of the patient, which motivates the need of early detection of these traumatic lesions.⁸

Several approaches for automatic detection to assist in visual analysis have been proposed, ranging from machine learning on whole images^9,10 to computer-assisted segmentations and morphological operations.^11,12 In recent years, machine learning gained broad attention in the field of medical imaging by detecting several abnormalities. Kumar et al.¹³ introduced a robust method by expanding multivariate generalized Gaussian distribution to a reproducing kernel Hilbert space for mixture modeling. According to the authors, this statistical learning method is robust compared with the previous studies that were highly sensitive to the outliers caused by several errors, including motion and imaging artifacts. The method was examined using retinopathy and gastric and esophageal cancer datasets. After comparing the results with seven other methods, the statistical learning method showed a higher accuracy.¹³ Another study published by Schlegl et al.¹⁴ used generative adversarial networks to find the markers that provide information regarding the progression and treatment of a certain disease. The method was examined on 8792 two-dimensional retina images for the detection of retinal fluid and other retinal lesions. This technique was able to identify several abnormalities.

An alternative technique is comparing findings with an average head CT that was first performed by Rorden et al.¹⁵ A template of subjects with a mean age of 65 years was intended for the detection of brain damage in stroke patients. Gillebert et al.¹⁶ used this template in comparing CT scans of stroke patients and showed detection of lesion boundaries with a sensitivity of 75%.

The aim of our study was to design and realize a simple and robust automatic detection method for intracranial hemorrhages smaller than 10 mm in diameter in patients with TBI, as these are most easily missed on the CT scan images in an emergency OR situation. In some cases, secondary injury occurs if these lesions are not detected promptly. Our main goal was to develop a method with the ability to detect the presence of any type and size of intracranial hemorrhage so that the emergency room physicians can call in a neuroradiologist, a resource not likely available 24/7 in smaller regional hospitals.

2.

Materials and Methods

3.

Results

As shown in Fig. 1(a), the skull of each patient in the control group was altered to fit the skull shape of the fixed image (white arrow), while the intracranial details differed. From these three-dimensional images, an average image was generated where each voxel in the average image represented the average of 30 voxels at that point [Fig. 1(b)].

Fig. 1

(a) The fixed image (white arrow) with 29 resulted images after image registration of the CTs of patients in the control group. All images are obtained from the same slice in the scan. The heads of the patients on the CT images have similar skull shapes with different detailed information in the brain. From these images, (b) an average image is generated.

The data of the average image of patients in the control group were compared with the CT images of TBI patients, and the hemorrhages were automatically visualized. According to the clinical notes, 67 lesions were detected where the type and the region of hemorrhage differed per patient (Table 2). The neuroradiologist and the radiologist in training missed the same two hemorrhages (sensitivity of 97%) that were present in the temporal lobe near the bone. The neuroradiologist found three false positives, and the radiologist in training detected two of these false positives. Figure 2 shows one of the true positive slices of the original CT of a TBI patient and the resulting image after automatic detection. The original CT image shows two small hemorrhages at the edge of the lateral ventricles. After applying the automatic detection technique, the hemorrhages are highlighted in green/yellow. Also, the edges of the skull were visualized as a high-intensity region. Since our main goal was detection of small intracranial hemorrhages, the skull was ignored in the visual assessment. However, to maintain the anatomical orientation in the resulted image, skull stripping was not performed. For one patient, our method showed a cerebral hemorrhagic contusion that was originally missed. In the resulted images of this TBI patient, the radiologists detected a minor asymmetry in the temporal lobe, which was not noted in the clinical data of the patient. After re-examining the original CT image of the patient, a cerebral contusion in the temporal lobe was detected. Figure 3(a) shows the original CT image of the patient with bilateral small subdural hemorrhages along Meckel’s cave (green arrows) and small initially undetected hemorrhagic contusion in the left temporal area (red arrow).

Table 2

Type and location of hemorrhages in the brain of TBI patients.

Fig. 2

(a) A slice of the original CT image of a TBI patient with two small regions of hemorrhage (green arrows). (b) After applying the automatic detection method on the original image, the two hemorrhages were easily distinguishable from the healthy tissue. However, high intensities were also detected at the edges of the skull.

Fig. 3

(a) A slice of the original CT image of a TBI patient with bilateral small subdural hemorrhages along Meckel’s cave (green arrows) and an originally missed cerebral hemorrhagic contusion (red arrow). (b) After applying the automatic detection method on the original image, the two subdural hemorrhages were detected, while the hemorrhagic contusion was partially detected. However, the hemorrhagic contusion was not noted in the clinical notes of the patient.

The final images of TBI patients also showed high-intensity regions along the cerebellar tentorium, pons, and around the lateral ventricles. Furthermore, the radiologists had no difficulties in discerning dense vessels and calcifications from hemorrhages in the obtained images.

4.

Discussion

In this pilot study, we mimicked the visual analysis of the neuroradiologists with an automatic detection method by easily identifying small cerebral hemorrhages in TBI patients using a computer-generated “average brain” from 30 control patients. From 67 lesions, the automatic detection method missed only two lesions that were present in the temporal fossa near the bone. The middle temporal fossa is highly variable in shape and size per individual. These variations were observed on the original CT scans of 30 patients in the control group. Since only 30 young patients were included in the control group, the average image will be affected by these individual differences. The shape of the fossa media is partially race dependent; therefore, development of several templates based on race may solve a part of the problem.

The automatic method detected a hemorrhage that was initially missed by a resident. After detecting the hemorrhage by our automatic method, two neuroradiologists and a radiologist in training critically re-examined the original CT image. The three physicians were able to confirm the presence of hemorrhage on the original CT image.

In most hospitals, the examination of the (neuro)radiologist and the resulting clinical notes are used as a gold standard for TBI patients. Since autopsy is usually not an option and our work is a retrospective pilot study, we used the visual analysis of our neuroradiologists as the gold standard.

Al-Ayyoub et al.⁹ examined CT images of 74 patients classified into normal, epidural, subdural, and intraparenchymal hemorrhages. They applied several techniques, such as image preprocessing, segmentation, region growing, feature extraction, and classification, resulting in an accuracy of 100% for the detection of hemorrhages. Shahangian and Pourghassem³⁰ performed automatic brain hemorrhage segmentation and classification algorithm based on weighted grayscale histogram feature in a hierarchical classification structure. Detection accuracies of epidural, subdural, and intracerebral hemorrhages were 96.15%, 94.87%, and 95.96%, respectively. Liu et al.¹⁰ used machine learning for the detection of cerebral hemorrhages in neurotrauma patients with a detection accuracy of 80%. The authors included mostly large hemorrhages that were not easily overlooked on the original CT images. In contrast to these studies, we focused on hemorrhages smaller than 10 mm in diameter. Comparing our pilot results to the results of these studies, our method generally falls in the same accuracy range. However, it is noteworthy that we have only included 9 CT scans in the TBI group and 30 CT scans in the control group. Additionally, our method did not require any segmentation of the skull, ventricles, or cerebrospinal fluid (CSF).

The original CT images of the brain are often initially read by a radiology resident or sometimes by a nonradiologist. Our automatic method may assist the radiology resident and other physicians in the detection of small traumatic hemorrhagic lesions. Since our method visualizes normal tissue with a blue color and other findings with other colors, the physicians may quickly know where to focus on the image. Automatic detection can save time by directly distinguishing lesions from healthy brain areas. Moreover, it may reduce the interobserver variation between radiologists.

In this study, we compared the CT scans of a relative young control group with TBI patients of different age groups. Since aging changes the brain, sometimes reducing it in size,³¹ it is more precise to compare patients with age-matched control groups. This may reduce misdetection and may cause less deformation of the average image to correctly align to the TBI images.

To initiate the image registration, a reference image was required to align the images. In this work, a CT image of a patient in the control group was selected as the fixed image. Since there was one fixed image and 30 moving images, the fixed image had a higher influence on the resulted images. An approach for eliminating the bias is a groupwise image registration where all images are aligned to a common space. Applying a groupwise registration, the information of all images will have an equal influence on the average image.^32–34 Recently, Elastix introduced the method of Huizinga et al.³² in its database for groupwise registration of MRI images. To our knowledge, using Elastix software, this approach has not been applied on CT images.

Several registration steps were applied to obtain a good trade-off between quality and speed. The goal was to create an average brain CT from several CT images to include the normal variability in brain tissue. After applying rigid and affine transformation, the details of the brain tissue in the resulted images looked nearly the same as the fixed image. Since the images were obtained from different patients, rigid transformation was not able to correctly align the images and to include the details of the moving images in the results. To perform the registration more precisely, nonrigid B-spline transformation was applied. The B-spline method uses a grid with automatically placed control points on the vertices of the grid squares that are overlaid on the image. The spacing of the control points determines the deformation. Large spacing describes a global deformation, whereas small spacing defines a local deformation. To match the small structures in the images, different spacing values were tested. As larger values of 30 and 40 mm resulted in a mismatched registration image, a spacing of 20 mm between two control points was preferred for the CT registration. The results of nonrigid transformation showed a good brain overlap that was confirmed by visual inspection. Using nonrigid transformation, the tissue density may change locally, especially in the transition regions from white to gray matter, CSF, and parenchyma. However, these density differences are still lower than the density of a hemorrhage. The densities of CSF and brain tissue are between 3 and 40 HU, and the density of blood is generally higher than 60 HU.^35,36 The automatic method is developed for detecting intracranial hemorrhages and will not be affected by other focal lesions, such as white matter lesions.

The resulted images of TBI patients showed higher intensity regions around the lateral ventricles. The reason for these high-intensity regions may be the large differences in contrast between the ventricles and the brain parenchyma. Since the shape and position of the ventricles differ per individual, voxel resampling to another brain may cause this misdetection. Furthermore, some TBI patients develop a midline shift that causes a displacement of the ventricles. This displacement can have an influence during the comparison of TBI scans to the control group and may cause the false positives in the regions around the ventricles. The neuroradiologist and radiologist in training did not identify the high-intensity regions around the ventricles as pathologic. It is, however, possible that other physicians may interpret these areas as false positives. A possible solution for this problem could be the segmentation of the ventricles before image registration.

Similar problematic areas were regions around the pons and/or cerebellar tentorium. A reason for these areas being highlighted could be artifacts on the original CT images that are difficult to distinguish from hemorrhage. Other reasons for the high intensities in the cerebellum could be subarachnoid blood or a consequence of our method. Subarachnoid blood is usually quickly diluted by the CSF, so it may not be present on a follow-up study. For these types of findings, prospective research with a follow-up MRI could provide additional information. Using this information, we can examine if these high-intensity regions are caused by our method or by other factors.

5.

Conclusion

We have introduced an automatic detection method for the detection of small traumatic brain hemorrhages in TBI patients using a computer-generated average CT. Our automatic detection method showed encouraging pilot results and a good correlation with the visual analysis of the neuroradiologists. The automatic comparison of individual CT scans with the computed average may assist the physicians in early detection of small hemorrhages.