A New Liver Segmentation Based on Digital Liver Portal Vein Ramification Using Computer-Assisted Surgery System: Exploring Artificial Intelligence

Abstract

A good understanding of liver anatomy is required for performing precise liver resection. However, the currently described methods of liver segmentation based on portal and hepatic veins are inconclusive. We proposed a system of liver segmentation based on previous reports and our data. Three-dimensional computed tomography software based on artificial intelligence was used to analyze the portal vein branching pattern in 759 patients. We analyzed four different types of liver segmentation and measured their respective segmental liver volumes. We classified four types of liver segmentation based on the right portal vein. Median segmental liver volumes were variable for the different types of segmentation. Our system of liver segmentation enables a better classification of individual patients into one of the different types, thus assisting in preoperative surgical planning. Segmental liver volume is useful for the preoperative evaluation of remnant liver volume.

1. Introduction

An accurate understanding of liver anatomy is important for surgical safety [1]. This is particularly relevant to the progression of modern surgery toward individualized treatment and the advent of partial hepatectomy and living liver transplantation technology [2, 3]. Initial liver segmentation studies were based on the cadaver liver specimen perfusion model and were limited by the number of specimens and the research techniques available at the time [4, 5, 6]. With the development of artificial intelligence, the development of modern imaging and digital medical research enabled the analysis of dimensional anatomical relationships and spatial vascular variations by three-dimensional (3D) visualization technology from all directions in a transparent and interactive manner [1, 3, 7, 8]. This has particularly helped with the performance of in vivo liver segmentation and liver volume measurement [9, 10]. In recent years, several studies have used digital imaging technology for liver segmentation, although these were often confined to liver lobe variations [11, 12, 13] and did not involve systematic research. None of the existing liver segmentation methods includes all possible variations in the liver anatomy. The portal vein branches are relatively consistent in the left hepatic lobe, which is divided into segments II, III, and IV. However, none of the existing single segmentation methods describes the different variations in the right liver. We describe a new liver segmentation system based on 3D reconstruction studies of digital liver models.

We used the 3D U-Net framework. In the field of machine learning, the U-Net is a successful encoder-decoder network that has received a lot of attention in recent years. Its encoder part works similarly to a traditional classification CNN in that it successively aggregates semantic information at the expense of reduced spatial information.

2. Materials and methods

A total of 759 patients without liver disease were enrolled in this study from July 2013 to November 2017. Upper abdominal contrast-enhanced computed tomography (CT) image data were collected for all patients. Patient selection criteria were as follows: (1) no liver lesions or other diseases affecting the portal vein arrangement; (2) availability of high-quality CT imaging that clearly displayed the portal vein up to its fourth-level branch; (3) no history of liver surgery; and (4) CT layer thickness less than 1 mm. This study was approved by the research ethics committee of the affiliated hospital of Qingdao University, and written informed consent was obtained from all parents.

All included patients underwent upper abdomen contrast-enhanced CT (Discovery HD 750; GE Healthcare, Milwaukee, WI, USA and Definition Flash; Siemens Healthcare, Forchheim, Germany). The scan parameters were set as follows: nonionic contrast agent (Iopromide 350 mg I/mL; Schering Ultravist, Berlin, Germany) was injected via the forearm elbow vein or the hand vein with a double-tube high-pressure syringe (Stellant; Medrad, Indianola, PA, USA). Approximately 1.5–2.0 mL/kg body weight of contrast was injected at a rate of 1.0–3.0 mL/s. For Definition Flash CT, the tube rotation time was 0.28 s, detector collimation was 2*64*0.6 mm, and pitch was 1.0. For Discovery HD 750 CT, the tube rotation time was 0.5 s, detector collimation was 64*0.625 mm, pitch was 0.984, and noise index was 10.

2.1. Image processing and 3D reconstruction based on artificial intelligence

DICOM data of the upper abdomen CT were uploaded into the Hisense Computer Aided Surgery System (Hisense CAS, version 2.1.3; Qingdao, China) for 3D reconstruction [3, 4, 5]. The following steps were performed: liver image extraction (liver segmentation was performed automatically through the artificial intelligence automatic adjustment of the window width and window level); extraction of intrahepatic vascular system (the scope of blood vessel formation was determined through the selection of intrahepatic vascular markers, followed by automatic extraction of intrahepatic vascular information); and integration (with integration of the liver and intrahepatic vascular system, 3D reconstruction was used to display the portal vein trunk, branch arrangement, and dominated region clearly from all directions in 3D).

2.2. Segmental volume measurement

By using the surgical simulation module of Hisense CAS, watershed analysis was performed based on the portal vein arrangement, radius, and supply area. The volumes of liver segments for types A, B, and C were calculated based on the fourth-level branch of the portal vein. Type D varied greatly and the number of each variant was small.

2.3. Basic principle for Dong’s liver segmentation

Based on statistical analysis, approximately 10% of cases were selected for pre-verification. Preliminary segmentation was performed and segmentation principles were proposed; these were verified by using the larger sample. The following basic principles for liver segmentation were developed based on the statistical analysis of preexperimental results obtained by the 3D reconstruction of the normal liver and vascular system of 120 humans.

1. The 3D model of the digitalized liver and vascular system was utilized to describe Dong’s liver segmentation based on the portal vein branch. The area supplied by the fourth-level portal vein is often considered the basic unit for precision liver resection. Therefore, the dominant area of the fourth-level portal vein was identified as the criterion for liver segmentation in Dong’s liver segmentation system.

2. The caudate lobe region of the liver (segment I) has a relatively special portal vein blood supply with large variations. Following the primary portal vein branch, three to six small blood vessel branches are derived directly from the left and right main branches of the secondary portal vein to supply the caudate lobe area. Generally, five to eight short hepatic veins allow for backflow of blood. During precision liver surgery, portal vein bleeding cannot be solved by blocking a third-level or fourth-level portal vein, as performed for other liver segments.

3. We defined the caudate lobe as segment I to respect tradition and to enable easy recall. The subsequent lobes, starting from the left lobe of the liver, were numbered segments II–IX in a clockwise direction.

2.3.1 I 3D reconstruction and intrahepatic vascular system Based on Artificial Intelligence

The 3D reconstruction of the liver and intrahepatic vascular system was performed for 759 patients of different ages. The reconstructed digital liver by use of machine learning appeared to have a clear structure (Figure 1). The spatial distribution and variation of the portal vein were observed in rotated directions, leading to the observation of the spatial anatomical relationships of the portal vein within the liver from different angles.

Figure 1.

(a–d) The three-dimensional (3D) reconstruction results of the liver, hepatic vein, and portal vein using Hisense CAS and integration.

2.3.2 II Dong’s liver segmentation

Type A is similar to Couinaud or Cho’s segmentation, with the liver containing eight segments (365 cases, 48.09%). Type B contains nine segments because of the three branches of the right-anterior portal vein (203 cases, 26.75%). Type C (76 cases, 10.01%) has two variations, type C-a, wherein the right-posterior portal vein is sector-shaped and the right-anterior portal vein is similar to that in type A, and type C-b, wherein the right-posterior portal vein is sector-shaped and the right-anterior portal vein is similar to that in type B. Type D contains special portal vein variations that need three-dimensional simulation to design individualized liver resection plans (115 cases, 15.15%).

2.3.2.1 Type A

Segment I is the caudate lobe, which is supplied by three to six small portal vein branches derived directly from the left and right main portal veins. Segments II and III are supplied by fourth-level portal vein branches derived from the superior and inferior outer aspects of the umbilical part of the left main portal branch. Segment IV is supplied by the fourth-level branch of the left portal vein.

The left lobe nomenclature, including segments I to IV, is similar for all four types (Figure 2). The right portal vein divides into the right anterior and the right posterior branches. The right anterior branch further divides into two main branches, the cephalic and caudal branches, or the ventral and dorsal branches, depending on the angle of their branching. The caudal or ventral branch supplies segment V, whereas the cephalic or dorsal branch supplies segment VIII (Figure 3a).

Figure 2.

(a–d) Liver segments I, II, III, and IV and their respective portal venous blood supply.

Figure 3.

(a) Branches of the right anterior portal vein and the liver segments supplied by them for type A. (b) Branches of the right posterior portal vein and the liver segments supplied by them for type A.

Segment VI is the area supplied by the fourth-level portal vein derived from the outer inferior aspect of the right liver following the third-level branch of the right portal vein branch. Segment VII is the area supplied by fourth-level portal vein derived from the superior outer aspect of the right liver following the third-level branch of the right portal vein branch (Figure 3b).

2.3.2.2 Type B

According to the portal vein branches and the dominant areas, the liver was divided into nine segments for type B. A total of 203 (26.75%) cases were type B. Segments I to IV are similar to that of type A.

The right portal vein divides into the right anterior and the right posterior branches. Segments V, VIII, and IX are the areas supplied by the three main branches of the right anterior branch: the caudal (portal branches of segment V, P5), dorsal (portal branches of segment VIII, P8), and ventral branches (portal branches of segment IX, P9), respectively (Figure 4a). Segments VI and VII are the areas supplied by the fourth-level portal vein derived from the outer inferior and superior aspects of the right liver, respectively, of the right posterior branch (Figure 4b).

Figure 4.

(a) Branches of the right anterior portal vein and the liver segments supplied by them for type B. (b) Branches of the right posterior portal vein and the liver segments supplied by them for type B.

2.3.2.3 Type C

The right posterior area of the liver is supplied by 5–11 sector-shaped branches of the portal vein branches that are derived from an arched main vessel (Figure 5). It is not possible for segments VI and VII to be resected individually with precision (but is possible for types A and B). The proportion of livers with type C is small but significant from the point of view of precision liver resections. Type C has two variations, type C-a (6.59%), wherein the right-posterior portal vein is sector-shaped and the right-anterior portal vein is similar to that of type A (P8), and type C-b (3.42%), wherein the right-posterior portal vein is sector-shaped and the right-anterior portal vein is similar to that of type B (P8 and P9).

Figure 5.

The right posterior portal vein and its branches supply the segments for type C.

2.3.2.4 Type D

Type D is a special group of variants that included 115 cases (15 15%). The various portal vein configurations identified were as follows. Among the common type, the portal vein divides into right and left branches and the right anterior portal vein branch is derived from the left portal vein (74 cases, 64 35%) (Figure 6a). The P6 portal vein is derived from the right anterior portal vein distal to the branching of the P7 portal vein from the right portal vein (19 cases, 16 52%) (Figure 6b). The portal vein trunk is trifurcation at the porta hepatis; it divides into the left, right anterior, and right posterior branches (8 cases, 6 96%) (Figure 6c). The right anterior portal vein is derived from the saccule of the left portal vein (4 cases, 3 48%) (Figure 6d). Approximately four to eight branches with similar thickness are derived from the right anterior portal vein supplying the right anterior liver (Figure 6e). The P2 and P3 branches share a common trunk (Figure 6f) that leads to several branches that supply liver segments II and III. The right anterior portal vein has a trunk that divides into several sector-shaped small branches.

Figure 6.

(a) The right anterior portal vein is derived from the left portal vein main trunk. (b) The P6 portal vein is derived from the right anterior portal vein. (c) The portal vein trunk has trifurcation at the porta hepatis and divides into the left, right anterior, and right posterior branches. (d) The right anterior portal vein is derived from the saccule of the left branch. (e) The right anterior lobe has a dominant supply from seven branches that are simultaneously derived from the right anterior portal vein. (f) The P2 and P3 branches of the left portal vein share a common trunk.

The different types of liver segmentations in the different sexes are shown in Table 1. There was no difference between the sexes in terms of the different types of liver segmentation (χ² = 2.823, p = 0.420) (Table 1). Similarly, there was no difference between the pediatric (3 months to 15 years) and adult groups (>15 years) (χ² = 1.095 and p = 0.778) (Table 2).

Segmentation	Sex		Total
	Male	Female
Type A	118 (46.27%)	247 (49.01%)	365 (48.09%)
Type B	77 (30.20%)	126 (25.00%)	203 (26.75%)
Type C	26 (10.20%)	50 (9.92%)	76 (10.01%)
Type D	34 (13.33%)	81 (16.07%)	115 (15.15%)
Total	255 (100%)	504 (100%)	759 (100%)

Table 1.

Distribution of liver segmentation in different sexes.

Male vs. female: χ² = 2.823, p = 0.420.

Segmentation	Age		Total
	3 months–15 years	>15 years
Type A	22 (50.00%)	343 (47.97%)	365 (48.09%)
Type B	9 (20.45%)	194 (27.13%)	203 (26.75%)
Type C	5 (11.36%)	71 (9.93%)	76 (10.01%)
Type D	8 (18.18%)	107 (14.97%)	115 (15.15%)
Total	44 (100%)	715 (100%)	759 (100%)

Table 2.

Distribution of liver segmentation in pediatric and adult groups.

Pediatric vs. adult age groups: χ² = 1.095, p = 0.778.

2.3.3 III Segmental volumes for types A, B, and C of Dong’s liver segmentation system

The volumes of each of the liver segments of the different types are presented in Tables 3–6. For type A, segments V and VIII account for 15.78% (±5.12) and 16.43% (±5.18) of the total liver volume, respectively. For type B, the volumes of segments V, VIII, and IX account for 10.36% (±3.72), 11.84% (±3.28), and 12.69% (±3.70), respectively. The volume of the right-posterior (RP) segment of type C was smaller than that of segments VI and VII of type A and type B (26.379% [±5.613] and 24.447 [±5.025], p < 0.01).

Liver segment classification	% of total liver volume (range)
	I	II	III	IV	V	VI	VII	VII
Type A (n = 365)	4.86 ± 1.89	10.16 ± 3.07	11.60 ± 3.52	14.27 ± 3.40	15.78 ± 5.12	11.65 ± 4.23	15.28 ± 4.83	16.43 ± 5.18

Table 3.

Volume ratio for each type A segment of Dong’s liver segmentation (%).

Liver segment classification	% of total liver volume (range)
	I	II	III	IV	V	VI	VII	VII	IX
Type B (n = 203)	4.79 ± 2.05	9.55 ± 3.02	11.70 ± 3.43	13.71 ± 3.49	10.36 ± 3.72	10.61 ± 4.17	14.77 ± 4.48	11.84 ± 3.28	12.69 ± 3.70

Table 4.

Volume ratio for each type B segment of Dong’s liver segmentation (%).

Liver segment classification	% of total liver volume (range)
	I	II	III	IV	V	RP	VII
Type C-a (n = 50)	4.99 ± 2.42	10.33 ± 3.31	12.32 ± 4.05	14.08 ± 3.06	16.88 ± 4.59	24.61 ± 4.70	16.69 ± 4.59

Table 5.

Volume ratio for each type C-a segment of Dong’s liver segmentation (%).

Liver segment classification	% of total liver volume (range)
	I	II	III	IV	V	RP	VII	IX
Type C-b (n = 26)	5.19 ± 1.98	9.70 ± 3.44	11.39 ± 3.66	14.03 ± 3.58	11.47 ± 3.82	24.14 ± 5.69	11.50 ± 3.52	12.50 ± 3.47

Table 6.

Volume ratio for each type C-b segment of Dong’s liver segmentation (%).

Accurate preoperative knowledge of the liver anatomy and volume is essential for performing safe liver resections [1, 10, 14, 15]. We proposed a liver segmentation system to enable better classification of the different types for individual patients to assist with their preoperative surgical planning. Segmental liver volume, which is useful for the preoperative evaluation of remnant liver volume, was also predicted.

Recently, many studies have proposed different methods of liver segmentation based on variations in the vascular anatomy of the liver [11, 12, 13]. Functional liver segmentation that included eight segments based on portal vein blood supply and hepatic venous drainage was most well-known and applied in clinical work [4]. However, the actual anatomical segmentation of the liver varied substantially in some cases.

Based on artificial intelligence, the development of imaging technology has enabled 3D reconstruction of the digital liver model in a CT DICOM file by using simulation software [16, 17]. The possibility of observing the anatomical relationship of the portal vein and hepatic vein in the liver from different angles allows for individualized evaluation of liver segmentation and subsequent surgical planning [18, 19, 20, 21, 22, 23]. Based on pediatric patients’ CT DICOM data, we developed software called Hisense CAS [17, 24, 25], which could accurately reconstruct the intrahepatic portal vein branches up to their fourth level. In the present study, we analyzed 759 digital livers; based on the variation of the fourth-level portal vein branch, we proposed Dong’s liver segmentation system. This system attempts to include all types of anatomical variations in the liver.

We found that the portal vein branches in the left hepatic lobe, which was divided into segments II, III, and IV, were relatively consistent. However, there were several variations in the right liver that cannot be described by a single segmentation method. Consequently, we classified them into four types: A, B, C, and D. For types A and B, segments VI and VII are supplied by a fourth-level portal vein derived from the outer, inferior, and superior aspects of the right posterior portal vein branches.

Couinaud divided cephalic segment VIII and caudal segment V based on the right anterior portal vein. This was disputed by a recent study that proposed that the right anterior divides into ventral and dorsal branches [11, 13]. Our study findings demonstrated that in some livers, the right anterior usually divides into two main branches, either cephalic and caudal or ventral and dorsal. We classified this as type A. Our observations suggest that a preoperative understanding of the angle of the portal vein branch is necessary for the right-anterior branch to avoid intraoperative injuries. Furthermore, the right anterior portal vein branch may also divide into three main branches, including the caudal (P5), cephalic-dorsal (P8), and cephalic-ventral (P9) branches. We classified this as type B.

The right posterior portal vein branch of type C is a single main branch with several small sector-shaped branches that supply the right-posterior lobe. These anatomical variations are important for segmental, subsegmental, and combined-segmental precision resections of the right liver.

A variety of special variations that could not be categorized into the first three types were included as type D. This strategy of grouping the special variants into one type may facilitate a full understanding of the complexity of liver anatomy. A higher proportion of such variance also supported the need for individualized precision surgery. To perform precision hepatectomy, liver segmentation should be performed based on individual liver models established using each patient’s imaging data preoperatively so that virtual surgery and remnant liver volume may be evaluated by using the 3D simulation software.

According to Kumon’s criteria, the caudate lobe (segment I) comprises three parts: the Spiegel lobe, the paracaval portion, and the caudate carina. Because of the uniqueness of the caudate lobe blood supply, the Spiegel lobe is supplied by one or two caudate lobe portal branches. Variations in the portal blood supply to the caudate lobe are very common. In our study, a considerable proportion of subjects were found to have two to five branches that were derived from the left and right portal veins supplying the caudate lobe. However, the actual number of such branches could not be accurately determined because of their tiny size and inadequate CT resolution. Because of these tiny vessels, we suggest that surgeons should demonstrate extra care during caudate lobe surgery.

We did not find any significant differences in sex and age using Dong’s liver segmentation. Because our data were from the Chinese population, the differences in the liver anatomy of people from various races and regions need further exploration.

The average volume of each segment of the different segmentation types can be used for predicting remnant liver volume to ensure safe anatomic hepatectomy. However, because of the large diversity in portal vein anatomy, it is our opinion that individualized volume measurements are critical for the safety of anatomic hepatectomy, especially in patients with large tumors, impaired liver function, or atypical portal vein branching.

Artificial intelligence technology has made significant breakthroughs and clinical applications in the field of precision surgery. The development of digital medicine has provided new perspectives regarding liver segmentation. We believe that Dong’s liver segmentation system and segmental liver volume will enable a better understanding of liver anatomy and will be useful for liver surgeons.