In radiotherapy treatment planning, target delineation is the premise of accurate and precise treatment delivery. Delineating treatment targets within DICOM files is complex, operator-dependent, and critical to the accurate delivery of conformal radiotherapy [11
]. These variabilities in target volume delineation can be a major primary source of inaccuracy of dose delivery and treatment errors [16
]. Consequently, efforts have been made to identify processes in the target delineation process amenable to improvement, such as multimodality image incorporation [8
], instructional modification [23
], visual atlas usage [11
], window-level adjustment [28
], auto-segmentation [29
], and software-assisted contouring [25
]. While specialized data entry mechanism for spatial data is common in other arenas (e.g., video games [31
] and virtual simulation workstations [32
]), there have been comparatively few efforts to modify ROI definition at the hardware level in radiotherapy.
Previously, ergonomic analysis by Kotani and Horii [2
] compared pen–tablet and mouse UIDs, demonstrating improved performance by pen–tablet UID on standardized repetitive computer drawing tasks (e.g., clicking, drag–dropping) and polygon tracing. The same series demonstrated EMG-detected reduction in muscular load to the flexor digitorum superficialis, extensor digitorum, and biceps brachii. Since, in its basic form, ROI outlining approximates polygon tracing as a fine motor task, evaluation of a pen–tablet UID was performed in this study.
The presented data suggest that specific aspects of the target volume ROI delineation process, across a range of tumor types and anatomic sites, are differentially impacted by the transition to a pen–tablet from a mouse–keyboard data entry system.
Both objectively measured active contouring time (e.g., excluding pauses >30 s in length) and subjectively estimated time contouring were reduced by a statistically detectable degree. However, the time savings did not appear to scale with total time required for each task and did not appear directly related to perceived case difficulty (Tables and ). The use of the pen–tablet UID was associated with a reduced number of corrective actions by individual users as well as reduced panning/scrolling functions, without reducing the frequency of contouring/drawing tasks recorded. Thus, it appears that the use of the pen–tablet results in ROIs that, while containing the same approximate number of drawn points, are less likely to be subsequently corrected than mouse-entered ROIs. The collected data also reveal no statistically detectable difference between pen–tablet and mouse–keyboard entry GTV ROI volume(s) for paired measurements (Table ). Likewise, intra-observer and inter-observer CN values were roughly comparable to those seen in a previous series [26
], suggesting that the ROI volumes designated were only minimally altered in a systematic manner based on data entry UID.
Subjectively, observers overwhelmingly preferred the pen–tablet entry system. Using collected data (Table ) as an ordinal scale, observers rated cases contoured with the pen–tablet as comparatively easier (paired Wilcoxon signed-rank test, p
0.04). Also, respondents were surprisingly accurate at time estimation, with a median difference between estimated and actual time spent delineating of <5 min. Consequently, not only was contouring with the pen objectively faster, but observers perceived the interface as faster.
While interest in target volume delineation has expanded in the conformal radiotherapy era, image perception evaluations [34
], workstation design [38
], and UID alteration [41
] have been less formally addressed than in diagnostic radiology. For example, Weiss et al. performed an evaluation comparing technologies used in diagnostic radiology. They compared QWERTY keyboard scroll wheel mouse to six different alternative UIDs (including five-button mouse, eight-button mouse, gyroscopic mouse, multimedia controller, handheld mouse/keyboard combination UID, and a gaming joystick) and found the standard mouse keyboard UID least favorable. In a similar study, Sherbondy et al. [41
] evaluated trackball, pen–tablet, jog–shuttle wheel, and mouse UIDs, finding the pen–tablet UID, in two distinct configurations, to perform faster than the mouse and trackball UIDs, respectively, at a simulated angiography localization task.
However, most literature on UID modification in diagnostic radiology literature focuses on scrolling, ROI localization/identification [41
], and annotation [43
] rather than target delineation, as in radiotherapy. Dowsett et al. [44
] reported a series of two gastro-/esophageal cancer cases, contoured by nine observers across several treatment planning systems (TPS). The treatment planning systems implemented distinct interface UIDs, including mouse–keyboard interface, light pen, and trackball input UIDs. In the Dowsett series, the TPS attached to each input UID was varied such that only a light pen and trackball could be compared on the same treatment planning system. Notably, in the present series, the use of the BB software substantially standardizes the contouring process such that all users were working with a common target delineation platform rather than on their typical institution-specific clinical workstations. Consequently, reflected task time estimates might be different with daily implementation on a commercially available TPS used regularly by the observer(s). More recently, Larsson et al. [45
] presented two abstracts comparing mouse–keyboard and the same pen–tablet model in this series using geometric shapes, as well as a lumbar vertebra [45
], as contoured by physicians, nurses, radiotherapists/dosimetrists, physicists, and administrative staff. In each case, the authors reported the pen–tablet interface to be faster at drawing tasks. Interestingly, they also observed that radiation oncologists were the slowest performers of ROI contouring tasks. The presented findings of this series, using more elaborate target volume-specific tasks designed to approximate clinical cases, correlate with those observed in simple geometric/anatomical shapes by Larsson et al. Work by Anderlind et al. [46
], using a pen UID with haptic feedback, furthermore suggests that the addition of tactile information might even further improve contouring efficiency.
Despite being the largest prospective UID comparison in therapeutic treatment planning, several caveats of this study are apparent; our sample size, though cumulatively robust, was limited to a selected subset of institutions, resulting in potential selection bias. Though multiple cases from distinct anatomic locations and different perceived task difficulty were used, the limited number of cases in each site limits broad applicability. Additionally, the use of non-parametric analyses owing to sample size considerations, the presence of several statistical outliers in the dataset (Figs. and ), and multiple non-Bonferroni-corrected comparisons, might make excessive generalization of the results erroneous. It is also noteworthy to consider these caveats in light of the large range of contouring times, the presence of several outliers (Fig. ), and the observations that times for contouring varied substantially for distinct cases (e.g., prostate case contouring was markedly faster than the head and neck case, regardless of device) and that individual users exhibited inter-observer differences in contouring speed severalfold greater than intra-observer UID-associated time improvement.
We did not survey participants for degree of familiarity nor preference with either UID before the study, which might also serve as an unidentified source of bias. By allowing users to contour cases in an order of their choosing, potential learning effects might be obfuscated. Thus, we sought to ascertain if contouring order was systematically by post hoc split-plot ANOVA analysis of active contouring time, UID, and both absolute (e.g., order of submitted contour sets for all four anatomic sites) and relative order (e.g., first or second contour submission for each organ site). In secondary analysis, neither absolute (F
0.46) nor relative order of completion (F
= 0.14) was found to be associated with a paired user difference in active contouring time.
Despite the stated limitations, our data demonstrate that modification of the input UID can detectably alter ROI delineation tasks. Specifically, in the tested cases, pen–tablet use was associated with decrement in active contouring time, perceived contouring time, perceived case difficulty, number of corrective actions, and number of panning functions while leaving ROI drawing and volumetric measures unaltered. In sum, the use of a pen–tablet device resulted in improved efficiency in ROI delineation tasks. These data suggest potentially appreciable savings in terms of physician time commitment. For instance, using the number of conformal/IMRT cases performed annually between 2006 and 2007, derived from site-specific (brain, prostate, head and neck, lung) numbers of conformal radiotherapy cases at the University of Texas Health Science Center at San Antonio and using time savings between input UIDs from the current dataset for each anatomic site, an estimated average ± SE of 1,335
916 min (22
15 h) of direct physician work time might have been saved annually from target volume delineation alone (e.g., not including normal structure/OAR ROI input or other time components of the treatment planning process) [47
] at a single participating site. Consequently, while time savings are obviously dependent on the number of cases, case mix, case complexity, and departmental size, among other factors, considerable institutional efficiency gains might potentially be realized from input UID optimization, especially given the relative time reimbursement costs for radiation oncologists [48
Finally, our data point to a need for a more developed analysis of image perception and human–computer interface evaluation specifically evaluating target volume delineation. It is imperative that future studies optimize technical parameters for accurate dose prescription. While this series did not directly evaluate clinical outcomes, it is possible that distinct display or human–computer interfaces might conceivably alter radiotherapy dose prescriptions sufficient to result in clinically meaningful sequelae, though such evaluation would require more robust numbers of cases.