Keywords
hemoperitoneum - CT scan - image processing - Hounsfield units (HU) - voxel segmentation - trauma care
Introduction
India grapples with the pressing issue of road accidents and the resulting fatalities. It is evidenced by a staggering 168,491 road crash fatalities in 2022, representing a 9% increase from the previous year and an 11.5% increase from 2019, according to the Ministry of Road Transport and Highways' report.[1] This alarming statistic underscores India's ongoing struggle with the critical challenge of road accidents and the loss of lives they entail. Notably, traumatic hemorrhage plays a substantial role in these mortality rates.[2]
[3]
[4] Hence, the timely and precise diagnosis within the critical golden hour (the critical time after injury when quick medical care can save lives) becomes pivotal in addressing this issue and saving lives.[5]
Understanding the complexities of hematomas and hemorrhages is paramount in managing injuries and trauma,[6] with particular importance placed on hemoperitoneum,[7] characterized by the accumulation of blood within the abdominal cavity, especially in cases of abdominal trauma.
Diverse tools and imaging techniques are employed for the identification and measurement of hemoperitoneum, encompassing manual inspection, palpation, computed tomography (CT), ultrasound imaging, magnetic resonance imaging, and computer-aided diagnosis systems. Among these, CT scans utilizing Hounsfield units (HUs) to gauge hemoperitoneum[8] are preferred for thoracoabdominal trauma detection due to their superior visualization capabilities.[9]
[10]
[11]
[12]
[13]
[14]
[15] They remain indispensable in diagnosing and managing internal bleeding, significantly contributing to treatment planning and informed decision-making processes.[16]
Existing Solutions and Challenges to Hemoperitoneum Diagnosis
Numerous classical image processing algorithms are available to tackle the challenges of achieving accurate segmentation and measuring hemoperitoneum volume. Nonetheless, the existing techniques and algorithms have limitations.
For instance, the Tada formula and its adaptations offer a basic estimate suitable for preliminary ellipsoid hemorrhage diagnosis, yet they fall short of accurately measuring sizable and asymmetrical hemorrhage.[17]
[18]
[19] The Tada formula's ellipsoid shape assumption has greater validity for the relatively homogeneous and limited cranial space. However, the abdominal cavity exhibits a wider variety of irregular forms and subject-specific variations, which limits its applicability for precise hemoperitoneum volume measurement. Region-based methods such as clustering, region-growing, contour-finding, and active contours rely on user expertise for precise segmentation.[20]
[21]
[22]
[23]
[24]
[25]
[26] Consequently, prevalent software packages incorporating classical image processing techniques, including region-growing algorithms, manual approaches, and spline methods, lean on explicit user inputs and offer restricted accuracy.[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34] Training and validation of statistical models for segmentation demand an extensive data set and manual CT scan segmentation—a task complicated by pixel intensity variations, border intricacies, high-contrast tissues, noise, and anomalies.[35]
[36]
[37]
[38]
Moreover, visually inspecting or manually assessing hemoperitoneum CT scans is vulnerable to intra- and interrater disparities, arbitrary imprecisions, and misinterpretations, all while consuming valuable time.[39]
[40]
[41]
[42] In light of these limitations, there is a clear need for more advanced and accurate approaches to address the complexities of hemoperitoneum diagnosis and segmentation.
Aim and Objective
The primary objective of this study is to retrospectively examine and compare the effectiveness of both novel and classical image processing algorithms in accurately segmenting hemoperitoneum areas in a group of 28 (n = 28) thoracoabdominal noncontrast CT images. We introduce HUVAO (Hounsfield Unit-based Volume quantification of Asymmetrical Objects), a novel algorithm utilizing region splitting and cluster selection guided by user inputs. We aim to evaluate how efficiently HUVAO segments and estimates hemoperitoneum volume. We evaluated its performance against visual estimations by two experienced radiologists and classical image processing techniques like Global Thresholding, Multi-Otsu Thresholding, and the Contour-Finding and Filling algorithm in estimating hemoperitoneum. Throughout our investigation, we aim to determine the accuracy of these algorithms in identifying and characterizing hemoperitoneum, examining their strengths and limitations, and exploring the potential for improving segmentation outcomes.
The study employs a slice-approach for hemoperitoneum quantification to achieve accurate volume measurement. The approach involves applying the selected segmentation algorithms to each slice of each CT image. The segmented regions obtained are then integrated to yield a comprehensive volume estimation of the hemoperitoneum. This approach has demonstrated its efficacy in delivering precise and reliable outcomes, irrespective of the shape of the hematoma, effectively minimizing shape-related errors and accounting for irregularities.[34]
Furthermore, the study delves into the intricacies of HU values within personalized imaging. By subjecting a novel algorithm to comparison with classical image processing methods, the study seeks to improve the accuracy and efficiency of hemoperitoneum segmentation. The ultimate aim, achieved through meticulous evaluation and exploration, is to identify the most effective algorithm capable of significantly enhancing diagnostic accuracy and improving patient outcomes, particularly in emergencies.
Methods
Data Description and Analysis
This retrospective study involved 28 patients who underwent non-contrast CT scans targeting the thoracic and abdominal regions. [Table 1] outlines the CT parameters encompassing the clinical data set utilized in this study.
Table 1
CT parameters of the study
|
KVP
|
120
|
|
X-ray tube current
|
176 mA-s
|
|
Exposure Time
|
320 ms
|
|
FOV
|
0.5100859375\0.5100859375
(0.5100859375 mm per pixel)
|
|
Slice thickness
|
1 mm
|
|
Slice gap
|
0.7 mm for the abdomen
0.5 mm for the thorax
|
|
Pitch
|
0.8
|
|
Matrix
|
512 (rows) * 512 (columns)
|
Abbreviations: CT, computed tomography; FOV, field of view; KVP, kilovoltage peak.
The CT scan data set was obtained in the Digital Imaging and Communications in Medicine (DICOM) file format. The MicroDicom v2022.1 viewer (64-bit) software was used to check it at first. This software is meant to handle pixel values and slice metadata in DICOM digital data.
Among the cohort of 28 patients under study, a demographic analysis reveals a gender distribution where merely two individuals were identified as female (7%), contrasting with the predominant male representation comprising the remaining 26 patients (93%), as depicted in [Fig. 1]. The age spectrum within this group ranged from 8 to 80 years, with a calculated mean age of 35.9 years and a median age of 34.5 years, as shown in [Fig. 2]. All subjects shared a commonality in the causes of their injuries, specifically road traffic accidents, and were presented at the hospital's emergency department within 24 to 48 hours post-trauma.
Fig. 1 Gender distribution among the 28 studied patients.
Fig. 2 Age distribution among the 28 studied patients.
Two experienced radiologists conducted visual estimations of the hemoperitoneum volume by thoroughly evaluating the CT scan images. Additionally, each radiologist independently reviewed the images and reported the observations regarding the presence and extent of hemoperitoneum. Any differences in their assessments were resolved through collaborative discussions.
Algorithms and Techniques Used
Due to the complex characteristics of the obtained clinical data, which include challenges like overlapping HU regions and the absence of well-defined HU ranges for accurate hemoperitoneum segmentation, our evaluation encompassed a novel algorithm and a variety of established classical image processing algorithms. The central aim was to segment and quantify hemoperitoneum within the dataset accurately. For a thorough comparison, we treated the mean of hemoperitoneum volume estimates from two radiologists not as a gold standard but as a practical benchmark, representing the “actual or reference value.” Visual estimates were used as a reference point for comparison alongside other quantitative methods, not considered the definitive reference value, to ensure a comprehensive evaluation.
The image processing algorithms involved are (1) HUVAO, (2) Global Thresholding, (3) Multi-Otsu Thresholding, and (4) Contour-Finding and Filling algorithm. Hemoperitoneum volume quantification adhered to the governing equation below, primarily grounded in the slice-approach for hematoma quantification[34]:
V
total
= (∑ pix) *(l*b*h) * 0.001 … (1)
Here, the total estimated volume of the region of interest (ROI) is referred to as “V
total,” while the CT scan voxel dimensions are denoted as (l, b, h). The term pix designates the total count of segmented pixels within a single CT slice along the XY axis. The inclusion of the number 0.001 in this formula enables conversion from cubic millimeters to milliliters.
The programming language used in this study is Python 3.8, and the libraries utilized include pydicom, scikit-image, numpy, and others.
HUVAO: the Novel Algorithm
This study introduces HUVAO, a novel algorithm that employs region-splitting and cluster-selection techniques guided by user input. HUVAO starts by determining the voxel dimensions of the CT scan (l, b, h) and collecting the user-provided predominant HU value (th) within the ROI. For each CT slice, pixel clustering is applied, grouping adjacent pixels with an absolute difference of 1 HU to unite closely related HU values, potentially indicative of the same tissue or structure. This threshold can be adjusted to accommodate different scenarios and user preferences.
After clustering, HUVAO receives a range of HU values (th ± a configurable constant value, e.g., [37 HU ± 1 HU]) centered around the initial user-provided HU value. Based on the average HU values of the clusters, the algorithm then identifies the clusters that fall within the predefined HU range. The minimum and maximum HU values for each selected cluster are recorded, and the algorithm establishes the final HU range by computing the respective mean of these values. Using the selected range, the algorithm performs a straightforward thresholding process to segment all slices within the ROI, calculating the total count of segmented foreground pixels (pix). This information is then used to determine the overall hemoperitoneum volume of one CT scan (V
total) following the formulation described in Eq. 1. The flowchart depicting HUVAO is presented in [Fig. 3].
Fig. 3 Segmentation and volume estimation by HUVAO (Hounsfield Unit-based Volume quantification of Asymmetrical Objects).
In the context of this clinical data set, global threshold values for hemoperitoneum segmentation were established through a comprehensive review of relevant literature sources.[43]
[44]
[45]
[46]
[47]
[48]
[49] These thresholds, set at 30 HU for the lower limit and 45 HU for the upper limit, were derived from the mode of HU values characterizing unclotted, non-contrast hemoperitoneum, covering both the lower and upper HU ranges. Additionally, when applying HUVAO to this data set, a threshold value (th) of 37.5 HU was utilized for hemoperitoneum segmentation. This choice was based on the calculated average between the previously mentioned threshold values of 30 and 45 HU, ensuring a balanced approach in the segmentation process.
Global Thresholding
Global Thresholding is a technique that divides an image into two segments using a single threshold value. Pixels brighter than this threshold are retained, while darker pixels are discarded.[50]
[51]
[52] This method effectively removes the background and facilitates object extraction. In the context of our clinical data set for hemoperitoneum segmentation, global thresholds of 30 and 45 were selected for this algorithm, as detailed in the previous section. The total hemoperitoneum volume of a single CT scan (V
total) was determined by utilizing the total count of segmented pixels (pix), as outlined in Eq. 1.
Multi-Otsu Thresholding
Multi-Otsu Thresholding is a technique that divides an image into multiple classes or segments (default is 3) by determining multiple threshold values based on pixel brightness, effectively categorizing different regions with varying brightness levels.[53]
[54]
[55] In the case of hemoperitoneum segmentation within our clinical dataset, experimentation determined that a minimum of seven classes were needed, and this configuration was adopted. The utilization of seven classes for segmenting the image histogram in each CT scan image led to identifying the optimal thresholding HU values for hemoperitoneum segmentation, precisely 31 (lower HU) and 43 (upper HU). The total number of segmented pixels (pix) was then used to find the total hemoperitoneum volume of one CT scan (V
total) as described in Eq. 1.
Contour-Finding and Filling
The Contour-Finding and Filling algorithm detects constant-value lines within an image but lacks information about the enclosed areas they outline.[24] In the case of CT scans with overlapping HU regions, the algorithm locates the contour coordinates of all the contours present in each CT slice. It involves considering regions with continuous boundaries. Subsequently, the algorithm employs a contour-filling process to fill all the identified contours based on their coordinate information. The algorithm utilizes the previously stated global thresholds (lower average HU = 30, upper average HU = 45) for selection of the filled contours in hemoperitoneum segmentation. The cumulative count of pixels within all the chosen and filled contours constituted the segmented pixels (pix). Subsequently, this total count of segmented pixels (pix) was used in the calculation of the overall hemoperitoneum volume for a single CT scan (V
total), as specified in Eq. 1. [Fig. 4] provides a visual depiction of the algorithm's workflow.
Fig. 4 Segmentation and volume estimation by Contour-Finding and Filling algorithm.
Results
In this section, we detail the clinical phase of the study, where we performed comprehensive segmentation and volumetric quantification tasks on a Windows 11 AMD Ryzen 7 machine with a 2.3-GHz CPU and 16 GB of RAM. The mean hemoperitoneum volumes, serving as the “actual volumes,” were determined through assessments by two radiologists. Several segmentation algorithms were used in the study, such as HUVAO, Global Thresholding, Multi-Otsu Thresholding, and Contour-Finding and Filling. [Figs. 5] and [6] display the volume estimates generated by each algorithm for all patients, with a deviation plot (the ratio of (“reference volume”/estimated volume) is used to quantify deviation) illustrating variations from the mean (reference) values provided by the clinicians in the cohort.
Fig. 5 Bar graph showing volume estimates obtained from the algorithms as compared with the clinicians' average estimates.
Fig. 6 Deviation plot of the algorithms from the clinicians' average estimates.
In terms of computational efficiency, the time complexities of the employed segmentation and quantification algorithms manifest as follows: HUVAO exhibited an average processing time of 10 minutes per CT scan, Global Thresholding took less than 5 seconds per CT scan, Multi-Otsu Thresholding demanded an average processing time of 7 minutes per CT scan, and the Contour-Finding and Filling algorithm had an average processing time of 31 minutes per CT scan.
Discussion
Significant challenges were encountered in the segmentation and volume estimation of the hemoperitoneum in this clinical dataset. These obstacles primarily arose from overlapping HU regions, limited localization precision, and the presence of small hemoperitoneum sizes, collectively complicating the accurate delineation and estimation processes.
Certain methods, like global thresholding and active contour methods, are great for high-contrast images[52] but struggle with overlapping HU regions. Multi-Otsu thresholding needs user inputs like the desired number of classes for image histogram separation[53]
[54]
[55] and does not work well for low contrast and overlapping region segmentation tasks. Clustering algorithms assume well-separated features, impractical for handling overlapping HU regions.[21] Despite being complex, contour-finding algorithm struggles to select suitable thresholds for overlapping areas, impacting accuracy.[24] Region-growing leads to oversegmentation and has specific prerequisites.[22]
[23] Compared with traditional image processing algorithms, HUVAO stands out for its user-friendliness due to minimal user inputs and its ability to deliver promising results, even when dealing with occasional misclassifications and quantification errors. For HUVAO to work at its best, it needs a dataset with clear visualizations, precise localization of ROIs, and distinct HU values linked to them. These features work together to accurately identify and isolate ROIs, making the visual assessment process more streamlined. [Fig. 7] showcases the separate segmentations of a randomly chosen CT slice with hemoperitoneum, demonstrating how all the algorithms were utilized for segmenting the clinical data.
Fig. 7 Hemoperitoneum segmentations: from top: left to right: Cropped and windowed image; Segmentation from an exhaustive search of Hounsfield unit (HU) values to minimize volume estimate error; Segmentation from HUVAO (Hounsfield Unit-based Volume quantification of Asymmetrical Objects); Segmentation from Global Thresholding; Segmentation from Multi-Otsu Thresholding; Segmentation from Contour-Finding and Filling algorithm.
In addressing the potential variability in hemoperitoneum detection based solely on HU values, an exhaustive analysis was conducted within our CT data set. The primary aim was to identify a specific HU value range for consistent and accurate hemoperitoneum segmentation, considering variations influenced by factors like hematocrit and blood age. Individual CT images were meticulously examined to identify common patterns in HU value ranges, supported by reference values provided by the two radiologists' visual estimations. The task involved determining minimum and maximum HU thresholds to minimize volume estimation errors and assessing the practicality of using overlapping HU values for segmentation. The findings emphasize the necessity of a comprehensive strategy for hemoperitoneum detection, acknowledging the limitations of relying solely on HU values in a clinical context. [Table 2] summarizes the outcomes of the extensive investigation, exploring lower HU thresholds from 32 to 39 and upper thresholds from 40 to 50, considering all combinations using simple thresholding techniques.
Table 2
Exhaustive searching of HU range with minimum HU: [30, 39] and maximum HU: [40, 50] producing the least volume estimate error
|
Patient. no.
|
Lower threshold
(in HU)
|
Upper threshold
(in HU)
|
Reference volume
(in mL)
|
Estimated volume
(in mL)
|
Deviation
(reference volume/estimated volume)
|
|
1
|
34
|
50
|
2500
|
2500
|
1.00
|
|
2
|
36
|
40
|
850
|
867
|
1.02
|
|
3
|
39
|
40
|
750
|
1330
|
1.77
|
|
4
|
39
|
40
|
175
|
742
|
4.24
|
|
5
|
39
|
40
|
850
|
699
|
0.82
|
|
6
|
36
|
40
|
1250
|
1301
|
1.04
|
|
7
|
33
|
44
|
1750
|
1753
|
1.00
|
|
8
|
38
|
42
|
1750
|
1730
|
0.99
|
|
9
|
36
|
42
|
2250
|
2249
|
1.00
|
|
10
|
37
|
40
|
1000
|
998
|
1.00
|
|
11
|
39
|
40
|
700
|
1252
|
1.79
|
|
12
|
38
|
40
|
1300
|
1341
|
1.03
|
|
13
|
34
|
40
|
1750
|
1740
|
0.99
|
|
14
|
33
|
44
|
2000
|
1999
|
1.00
|
|
15
|
39
|
40
|
750
|
649
|
0.87
|
|
16
|
39
|
44
|
900
|
916
|
1.02
|
|
17
|
38
|
42
|
1000
|
982
|
0.98
|
|
18
|
39
|
41
|
1350
|
1202
|
0.89
|
|
19
|
39
|
42
|
1100
|
1099
|
1.00
|
|
20
|
38
|
40
|
800
|
822
|
1.03
|
|
21
|
38
|
40
|
750
|
809
|
1.08
|
|
22
|
39
|
42
|
1750
|
1626
|
0.93
|
|
23
|
39
|
42
|
900
|
864
|
0.96
|
|
24
|
37
|
40
|
1000
|
1028
|
1.03
|
|
25
|
39
|
40
|
1000
|
839
|
0.84
|
|
26
|
39
|
40
|
350
|
341
|
0.97
|
|
27
|
35
|
40
|
2000
|
1985
|
0.99
|
|
28
|
39
|
40
|
225
|
329
|
1.46
|
Abbreviation: HU, Hounsfield unit.
A thorough analysis of the data presented in [Table 2] reveals a lack of consistent and accurate estimation of hemoperitoneum volumes through thresholding across the 28 CT scans. The variation in HU thresholds employed for hemoperitoneum segmentation among the scans highlights a lack of uniformity within the data set. This finding emphasizes the inherent difficulties in obtaining precise volume estimates for hemoperitoneum segmentation.
It is essential to note that depending solely on HU values might not be ideal for scenarios involving sensitive medical information. Various factors, such as tissue composition, density, and water content, can impact the HU values of different tissues. Additionally, specific tissues like the liver and spleen can share similar HU values with the hemoperitoneum, making differentiation more challenging. These findings highlight the limitations of image processing algorithms in consistently and accurately segmenting structures in CT scans with overlapping HU regions.
Looking ahead, further research toward refining HUVAO and addressing its limitations holds promise. Exploring strategies to enhance the differentiation of tissues with overlapping HU values and refining the algorithm's performance for smaller hemoperitoneum sizes could lead to improved outcomes. Additionally, investigating the integration of HUVAO with advanced technologies, such as deep learning approaches, might offer solutions to the challenges posed by overlapping HU regions. This research direction has the potential to advance the field of medical image processing and contribute to more accurate and reliable diagnoses in complex clinical scenarios.
Conclusion
The subsequent pathways for future investigation are identified based on the findings of our present study, the methodologies implemented, and the resultant outcomes:
-
(1) The potential approach is to obtain binary annotations for slices containing hemoperitoneum and train a convolutional neural network (CNN) using the annotated data set. The CNN can utilize Class Activation Maps to visually highlight predicted regions within each slice, aiding in further analysis and interpretation.
-
(2) Another approach involves collaborating with a radiologist to create voxel masks. Neural networks will then be trained to produce probability maps for each slice, simplifying automated pixel identification within the designated voxels. This method has the potential to automate voxel labeling.
Contribution to the Knowledge Domain
Our research contributes to and advances our understanding of medical image processing in the context of hemoperitoneum identification and segmentation using non-contrast CT scans. By evaluating a wide array of image processing algorithms, both classical and novel, we shed light on challenges in accurately segmenting structures, especially with overlapping HU regions. We emphasize that relying solely on HU values for blood detection in such cases might yield inconsistent results due to inherent variability influenced by factors like hematocrit, blood age, and contrast agents. Integrating HU values with other modalities and clinical data emerges as essential for precise diagnostics. Our comprehensive exploration reveals the complexities of achieving accurate segmentation within overlapping HU regions. This insight encourages the exploration of advanced techniques like deep learning and combining other methodologies. This knowledge informs the development of improved segmentation modules and promises to enhance diagnostic accuracy and patient care.
