We selected the shoulder for this experiment given its anatomical complexity and because it is considered one of the most difficult joints for medical students to understand in human anatomy. The Department of Anatomy and Developmental Neurobiology, University of Tokushima carefully examined anatomical accuracy, such as the relative spatial relationship of each structure, at every development stage to ensure that our interactive 3DCG models would be of high quality and accurate (see Multimedia Appendix 1).

The process of model creation is not trivial. We built the models in LightWave 3D (NewTek, Inc., San Antonio, TX, USA) and exported them as object files to Blender (an open source 3D program; blender.org, Amsterdam, the Netherlands). Blender was used to generate clean U and V space texture maps for the models. The models were then sent back to LightWave as object files. We created the textures in Photoshop (Adobe Systems Incorporated, San Jose, CA, USA) from photo references, applied them to U and V space maps in LightWave 3D, then exported them in Filmbox format and imported them into Unity3D [22]. Shoulder movements were added in LightWave and exported as Filmbox animation data. Once the model and animation data were imported into Unity3D, we wrote scripts to allow interaction with the models. Unity’s workflow allows swapping of models if changes have been made in LightWave 3D. The initial animation of shoulder movement was quite rigid, so we used a motion capture system to achieve more natural motion.

We used the Vicon MX motion capture system (Vicon Motion Systems, Oxford, United Kingdom) at Tokushima University Hospital [23]. Motion capture data were collected at 150 Hz using a near-infrared (780 nm) passive 8-camera system (Vicon MX T20; Vicon Motion Systems). A 3D position sensor captured the reflected rays of 9.5 mm diameter reflective markers. Nexus 1.4.1 (Vicon Motion Systems) software recorded 3D positions of the markers and extracted vector data. We took screenshots of the motion capture vector data and then moved the shoulder to match those screenshots for each part of the motion. We correlated the skeleton to the visual data, taking screenshots of data over time. This process improved the smoothness of shoulder movements.

The menu for the 3DCG interactive manipulation tools on the left side of the screen has two functional components: one for anatomy and the other for shoulder movements, with labeling ability in English or Japanese. Figure 1 shows the 3DCG of the area surrounding the shoulder and the terminology of each part on the upper right corner. Figure 2 shows the movements of the shoulder bones and upper extremity. Several important movements are available, including elevation, depression, retraction, protraction, flexion, extension, vertical abduction, vertical adduction, horizontal abduction, and horizontal adduction. Both views allow the user to zoom in and out, and to focus on a region of interest. Users are also able to quickly rotate or move to a specific angle such as anterior, lateral, and posterior. To study specific anatomical regions, the tool enables users to select single or multiple objects and hide them or make them semitransparent. We also developed an original textbook for this study. To provide a suitable condition for comparison, we replaced the black and white figures traditionally used in lectures at HSUM, which were drawn from a standard anatomical viewpoint, with a new textbook based on 3DCG images selected from the interactive 3DCG system. We reproduced 34 images (anterior, lateral, posterior, and other views with appropriate angle and magnitude for evaluation) from the 3DCG shoulder models in gif format for the textbook and added appropriate text to explain the images.

We used Wideband InterNetworking engineering test and Demonstration Satellite for the communication system, which was jointly developed by the Japan Aerospace Exploration Agency and the National Institute of Information and Communications Technology of Japan. A small antenna 1.2 m in diameter receives up to 155 Mbps of data and transmits up to 6 Mbps, while an antenna approximately 5 m in diameter enables 2-way communication up to 1.2 Gbps [24]. For our study, the Japan Aerospace Exploration Agency allocated reception bands from 1.31 Mbps to 1.38 Mbps for the uplink connection, and from 15.0 Mbps to 20.7 Mbps for the downlink connection. The link was encrypted using an Internet Protocol Security virtual private network based on a Cisco ASA 5505 router (Cisco Systems Inc., San Jose, CA, USA) provided by a joint research project with Mitsubishi Electric Information Network Corporation (Figure 3).

Round-trip time is the signal delay between the University of Tokushima and HSUM via the satellite network and the Internet. The round-trip time was between 1499.3 ms and 643.8 ms, with an average of 729.6 ms. The congestion window is a Transmission Control Protocol parameter that regulates the send window. The congestion window of Windows XP ranges from 16 KB in default to 64 KB. Therefore, the maximum throughput available in the Transmission Control Protocol is 0.795 Mbps, which is used for Hypertext Transfer Protocol. However, the actual throughput value was 0.384 Mbps or less.

All activity associated with the course was hosted on a Linux-based server running the Apache Web server (Apache Software Foundation, Los Angeles, CA, USA), the PostgreSQL database server (PostgreSQL Global Development Group, http://www.postgresql.org/), and the CentOS operating system (CentOS Project, http://www.centos.org/). Our proprietary learning management system (LMS) allowed us to create a course website with a unique log-in password and ID for each student. We developed the interactive 3DCG model, textbook, and questionnaires, and embedded them into this LMS. These components were used for the related experimental section. We applied Java Web applications for the system and adopted Unity3D for the 3DCG container, which is an integrated 3D platform for 3D games and interactive content on the Web.

Figure 4 shows the study design. All participants who had taken anatomy classes and finished cadaver dissection 3 months previously received a brief introduction to the LMS before the experiment. The study was conducted over a 1-week period. We divided participants into a 3DCG group and a textbook-only (control) group. The control and 3DCG groups were also given instructions on how to use the textbook, and the 3DCG group was given additional training to manipulate the interactive 3DCG model. Each participant received a 1-hour training session. Finally, all participants completed a questionnaire on the LMS.

Statistical analysis was performed using SPSS (version 16.0 for Windows; IBM Japan Inc., Tokyo, Japan). We conducted both the independent-samples t test and the Mann-Whitney U test to compare the two groups, with P < .05 defined as statistically significant. The internal consistency coefficient (Cronbach alpha) was calculated for both groups.

A total of 100 first-year medical students (27 men and 73 women; Table 1) from HSUM volunteered, gave informed consent to participate in the study, and were randomly assigned to either the 3DCG group (n = 50) or the control group (n = 50). The mean age of participants was 18.1 (SD 1.1) years (men: 17.9 (SD 0.6) years, women: 18.1 (SD 1.1); range 16–25 years). Participants were freshmen at HSUM and were instructed in practical computer skills for half a semester (64 hours).

The questionnaire had 24 items grouped into four categories: content (3 items), teaching methods (6 items), Web (3 items), and 3DCG model interface (12 items). Each item was scored based on a 5-point Likert scale: strongly agree (score 5), agree (score 4), neutral (score 3), disagree (score 2), and strongly disagree (score 1).

Table 2 shows that mean scores ranged from 3.77 to 4.14. For each item, 19%–38% of participants responded strongly agree, 40%–59% responded agree, 13%–26% responded neutral, 0%–9% responded disagree, and 0%–1% responded strongly disagree. As a result, 64%–86% of participants responded either strongly agree or agree to each item.

Table 3 shows that mean scores for the 3DCG module range from 3.96 to 4.62. For each item, 22%–66% responded strongly agree, 28%–70% responded agree, 2%–26% responded neutral, and 0%–1% responded disagree. None of the participants responded strongly disagree to the statements. Overall, 74%–98% of the participants responded either strongly agree or agree. The reliability of the entire questionnaire was acceptable (Cronbach alpha = .902).

Table 4 and Figure 5 present the mean (SD) scores. Table 4 compares scores between the 3DCG and control groups for each item. The 3DCG group scores ranged from 3.90 (SD 0.70) to 4.52 (SD 0.58), and control group scores ranged from 3.40 (SD 1.03) to 3.96 (SD 0.81). Differences in learning motivation scores were the largest, with the 3DCG group giving an average score 0.8 higher than the control group. Participants’ comments on the advantages of using the interactive 3DCG model included “very interesting,” ”realistic,” “saved time,” “more understandable,” “amazing movements,” and “clarity.” There were also some comments regarding disadvantages, such as the necessity for “muscle movement,” “more detailed anatomy,” and “larger screen size,” which would require compilation to change the frame size.

Combined scores for the three categories are shown in Figure 5. There was a significant difference between the 3DCG and control groups for content (4.26 (SD 0.69) vs 3.85 (SD 0.68); P = .001) and teaching methods (4.33 (SD 0.65) vs 3.74 (SD 0.79); P < .001). No significant difference was found between the groups in the Web category. The mean score for male participants was higher than that of female participants, except for Q11 (Figure 6). There was no significant difference between men and women for Q1–Q8 and Q10–Q12, but the difference for Q9 was significant.