Artificial Intelligence in Radiation Treatment Planning: A Survey-Based Observational Study From A Cancer Centre in Nigeria

Introduction

Cancer constitutes a significant global health burden, particularly in Low and Middle-income countries (LMICs) such as Nigeria, located in Sub-Saharan Africa (SSA). ¹ These nations face disproportionately high rates of cancer incidence and mortality that are increasing at a faster rate than in developed countries. ² This trend is largely due to limited awareness of cancer types and risk factors, alongside inadequate resources for patient support. Owing to a lack of cancer awareness, many patients are diagnosed with advanced-stage disease, ³ significantly increasing the demand for effective and advanced treatment options.

There are eight government-funded radiotherapy cancer centers in Nigeria for a total population of 223.8 million, which puts significant pressure on the scarce available resources. ⁴ Nigeria requires a minimum of 280 RT machines for the increasing number of cancer cases. ⁴ The radiotherapy pathway demands a significant investment of time and effort, requiring hundreds of hours dedicated by specialist healthcare professionals for each patient. To enhance patient throughput, it is crucial to reduce treatment and treatment planning times. One effective strategy is using hypo-fractionated radiation therapy, which delivers larger doses of radiation over fewer sessions, thereby shortening the treatment duration without compromising effectiveness. ⁵ Additionally, advancements in Artificial Intelligence (AI) can streamline various processes, such as optimizing treatment plans and improving patient scheduling. ⁶ By adopting these methods, healthcare providers can efficiently reduce wait times and enhance overall patient care. During the radiation treatment planning phase, oncologists, medical physicists, and dosimetrists collaborate to create the radiation plan to target the tumor while minimizing exposure to healthy tissues and organs. Precise contouring of tumors and nearby organs at risk (OARs) is crucial for effective monitoring of radiation doses.

Traditionally, this relies on providers manually tracing structures in Computed Tomography images, which can vary by provider and take over an hour. AI-driven auto-contouring significantly speeds up this process, though providers must still ensure the accuracy of the contours, especially with non-standard patient biology. ⁷

AI tools assist in cancer screening, clinical risk prediction, and radiotherapy planning. ⁸ AI is moving beyond algorithm development and integrating into clinical practice, transforming cancer research with improved risk prediction, precise treatments, and enhanced evidence generation. ⁹ Educating clinicians about AI systems is critical during this pivotal phase of AI integration into cancer care. Many healthcare providers are understandably cautious about this technology. Effective AI, machine learning, and deep learning require algorithms to be trained on existing data. ¹⁰ Furthermore, concerns regarding the ethical and legal implications of selecting patient data for training underscore the necessity of utilizing diverse data responsibly and without exploitation. ¹⁰ This study aims to evaluate the integration of AI into treatment planning among radiation oncologists and medical physicists at a tertiary cancer centre in Southwest Nigeria over a 6-month period.

Methodology

Statistical Analysis

Quantitative data were exported to Microsoft Excel for cleaning and descriptive summary (counts, percentages, medians, interquartile ranges). Paired comparisons of AI versus manual contouring times (oncologists reporting both methods) were performed using two-sided Wilcoxon signed-rank tests. Effect sizes (rank-biserial) and exact P-values are reported; where appropriate bootstrap 95% confidence intervals were calculated for median differences. Qualitative free-text responses were collated and summarised narratively. Given the exploratory nature and limited sample, inferential results are presented as exploratory evidence of effect rather than confirmatory hypothesis testing. The completed Checklist for Artificial Intelligence in Medical Imaging (CLAIM) is provided as Supplementary File and was used to guide reporting of this manuscript.¹⁷

Results:

Out of 31 respondents analyzed in this study, 64.5% (20) were Oncology consultants and Residents while 35.5% (11) were medical Physicists.

Table 1 above describes the demographic characteristics of the survey respondents. In total 31 staff completed the questionnaire; the cohort was predominantly clinical, comprising 20 oncologists and 11 medical physicists. Notably, age and sex were not collected in this survey and therefore are not represented in Table 1; this limits our ability to examine demographic confounders and is acknowledged as a study limitation.

OPEN IN VIEWER

Table 1.

Demographic Summary.

Demographic variable	Value
Total respondents (N)	20
Median age (IQR)	Not collected
Sex (M/F/Other)	Not collected
Occupation (counts)	Oncologist: 20; Medical Physicist: 11

At the time of the survey, 33% of all respondents had less than 6 months of experience with the AI contouring tool, followed by an even split of six months to a year and one to two years with 27% of respondents each. Finally, 13% of participants had over 2 years of experience with the AI tool.

When analyzing responses by occupation, it was noted that among medical physicists, six participants had less than one year of experience, one had between one and two years and three had over two years of experience. Among oncologists, seven participants had utilized the AI tools for less than six months, four had employed the tools for up to a year, and 7 had been using it for between one and two years. Only one oncologist had experience using the AI contouring tools for over 2 years.

Time it Takes Oncologists to Contour Using AI Tools Vs Manual Contours

Across the four analyzed cancer sites, oncologists consistently reported that the implementation of the AI contouring tool reduced their overall contouring time when compared to using manual contour/regular contouring. In contouring for breast cancer, 42.5% of oncologists spent less time (15 to 30 min) using the AI tools, while a significant number 26.8% required 30 to 45 min, 27.3% spent 45 to 1 h to complete it using manual contour.

Regarding cervical cancer, the respondents 37.8% finished contouring in less than 15 min and 37.0% in 15 to 30 min using the AI tools whereas with manual contour a significant number of respondents spend 42.5% spend 30 to 45 min to finish contouring. Continuing this trend, prostate cancer contouring took up mostly less than 30 min among 72.2% of oncologists using the AI tools, whereas manual contour necessitated 30 to 45 min or longer.

In the case of head and neck cancers, most (41.5%) oncologists spend 30 to 45 min for AI-assisted contouring with manual contour extending 45 min to an hour or more.

When assessing the time it takes for medical physicists to complete treatment planning using AI versus manual contour, breast cancer planning times have moderately improved with AI. Approximately 60% of respondents reported finishing their plans within 30 to 45 min using AI, while 30% completed their plans in the same time frame when using manual contour. Another 30% spent over one hour on their plans with manual contour. (Figure 4)

OPEN IN VIEWER

Figure 1.

Occupational Split of Respondents. One Figure in the Graph is Representative of one Respondent.

OPEN IN VIEWER

Figure 2.

Length of Experience with Using AI Tools for Contouring (top). Medical Physicist-Specific Experience is on the Bottom Left While Oncologist-Specific Experience is on the Bottom Right.

OPEN IN VIEWER

Figure 3.

(a) Time it Takes Oncologists to Contour Using AI Tools Versus Manual Contour (Breast Cancer). (b) Time it Takes Oncologists to Contour Using AI Tools Versus Manual Contour (Cervical Cancer). (c) Time it Takes Oncologists to Contour Using AI Tools Versus Manual Contour (Prostate Cancer). (d) Time it Takes Oncologists to Contour Using AI Tools Versus Manual Contour (Head and Neck Cancer).

OPEN IN VIEWER

Figure 4.

Time it Takes Medical Physicists to Complete Treatment Planning Using Manual Contour Tool Versus AI Tool.

The planning times for cervical and prostate cancer did not show significant differences. For cervical cancer, 50% of the respondents spent less time using AI, while the other 50% spent less time using MC. In the case of prostate cancer, approximately 40% of respondents utilized 15-30 min of MC, and another 40% spent 30-45 min using MC. Additionally, there was no difference in planning times for head and neck cancer when comparing AI and MC. (Figure 4).

Figure 5 illustrates the impact of AI tools on workflow and user experience as reported by oncologists and medical physicists. Most respondents agreed that the AI tool seamlessly integrated with their existing workflow and has improved the efficiency of radiation planning processes. There was notable hesitance among respondents to fully trust or rely on the AI tools within clinical practice, connecting back to the necessity for manual quality control and quality assurance in using the tools. Furthermore, responses varied significantly regarding the perceived improvements in accuracy following the implementation of AI tools into the clinical workflow. Only 19% of respondents strongly agreed that there were accuracy enhancements, while 39% somewhat agreed, and 41% of respondents either perceived no improvement or a decrease in accuracy. Satisfaction with the training and support for using the AI tool showed a significant range, with 3% of respondents indicating strong dissatisfaction.

OPEN IN VIEWER

Figure 5.

Impact of AI Tools on Clinical Workflow. Answered on a 5 Point Scale from Strongly Agree to Strongly Disagree. Answered by Both Oncologists and Physicists. Questions Marked with an Asterisk are Responses out of 30 Responses Instead of 31.

Impact of the AI tools on user experience are displayed in Figure 6. Respondents largely reported that AI tools have positively influenced their practice in the radiation oncology unit, and to a slightly lesser extent, they would recommend these tools to other professionals in the field. While many respondents felt that the AI tools had significantly influenced their overall confidence in delivering optimal patient care, or that it had improved their ability to develop personalized treatment plans, 6% felt that it hadn’t at all. Finally, most respondents felt the AI tools had increased their confidence in radiation planning of certain cancers.

OPEN IN VIEWER

Figure 6.

Impact of AI Tools on User Experience. Answered on a 5 Point Scale from Strongly Agree to Strongly Disagree. Answered by Physicists and Oncologists.

Primary benefits and limitations of using the AI tools are displayed in Figure 7. Most respondents felt that increased workflow efficiency (81%) and improved accuracy (58%) were the two primary benefits of integrating the AI tools into their workflow. Primary limitations included software licensing and technical know-how followed by network availability in 65%, 61% and 35% respectively. Cost of purchasing the AI tool, time, accuracy and hardware (computer) availability were other reported limitations.

OPEN IN VIEWER

Figure 7.

Primary Benefits (L) and Limitations (R) of Using the AI Tools. Answered by all Respondents. Each Respondent Could Answer with More Than one Answer.

In assessing AI tool accuracy, 35% of respondents felt that it was about 80% accurate (a score of 8 out of 10), displayed in Figure 8A. The medical physicists were probed farther, asking how often they have to manually verify or correct an AI plan and how effectively the AI tool can predict and manage OAR doses in their plans (Figure 8B and C). They were nearly equally split on how frequently they had to manually verify these plans, with 20% rarely having to, 30% sometimes having to and 10% frequently having to predict and manage OAR tolerances. In predicting and managing OAR tolerances, the majority of medical physicists believed the AI tool was highly effective, with none indicating it was ineffective.

OPEN IN VIEWER

Figure 8.

(A): Answers from Both Oncologists and Medical Physicists on AI Tool Accuracy When Assisting with Treatment Planning and Delivery. (B) Answers from Medical Physicists ONLY on Instances of Manual Verification/Correction. (C) Answers from Medical Physicists ONLY on Instances of Predicting OAR Tolerance Doses.

Statistical Analysis of Difference of Median Times Between AI Contouring Times

This panel presents paired AI-assisted and manual contouring times reported by oncologists (n = 19). The boxplots show that median contouring time with AI is approximately half that of manual contouring (=24.5 vs 46.3 min). Individual paired lines almost universally slope upward, indicating that for nearly all oncologists AI reduced time relative to manual methods. The Wilcoxon signed-rank test confirms this reduction is statistically significant (P < .001), and the rank-biserial effect size indicates a large practical effect.

This plot shows paired times for cervical cancer contours among oncologists (n = 19). Median AI-assisted contouring time is substantially lower than manual contouring (=22.5 vs 42.9 min). The paired observations indicate consistent time savings across respondents. Inferential testing (Wilcoxon) shows a statistically significant reduction (P < .001) with a large effect size, supporting a meaningful operational benefit.

Paired prostate contouring times (oncologists, n = 19) demonstrate clear reductions in median time with AI (=24.2 vs 42.9 min). Paired lines display some variability in magnitude of saving between respondents, but the overall pattern is a consistent reduction; the Wilcoxon test yields a highly significant P-value (P ≈ .001), indicating the change is unlikely due to chance.

Head & neck contours take longer overall than other sites, but oncologists still reported meaningful reductions when using AI (median =35.0 vs 55.0 min). Paired lines reflect more variability versus other sites, consistent with the greater complexity of head & neck anatomy, but the median difference remains significant (P < .002) with a large effect size, indicating AI confers clinically relevant time savings even for complex cases.

Discussion

Artificial intelligence (AI) is moving beyond the creation of algorithms and into common clinical practice in radiotherapy. It was a prospective, survey-based observational study in which we tested the impact of a commercial AI auto-contouring solution on the workflow of the staff in a tertiary cancer centre in Nigeria. The Results provide the demographics and baseline experience of the participants (Respondent occupation and experience are summarised in Figures 1 and 2, respectively). The significance of these baseline data is that the time-savings and attitudes observed are observed in the light of rather limited previous exposure to AI in many respondents.

In general, AI-assisted auto-contouring was linked to uniform time savings in clinician contouring time between tumour sites. Figure 3a-d displays oncologist contouring times (AI vs manual) of each tumour site and the paired-time comparisons (boxplots) are in Figure 9a-d; the paired-time comparisons indicate statistically and clinically significant median reductions of various common sites. Prostate and breast cases showed the largest relative benefit, which is also consistent with the relatively good OAR anatomy of those areas; the head and neck cases showed smaller relative gains, which could be due to the increased complexity of the anatomy that still needs to be manually refined.^11,12

OPEN IN VIEWER

Figure 9.

(a) Breast Cancer (AI vs Manual Contouring Time). (b) Cervical Cancer (AI vs Manual Contouring Time). (c) Prostate (AI vs Manual Contouring Time). (d) Head & Neck (AI vs Manual Contouring Time).

The planning time of medical physicists was also inconsistent with the oncologist contouring time as a measure of task heterogeneity, and not a malfunction of the AI tool. The times of physicist planning are presented in Figure 4. AutoContour generated OAR structure sets only in our workflow; target volumes (GTV/CTV/PTV) were marked and approved by clinicians. In this way, the timing of dose-optimisation and plan-completion activities downstream of contouring is found in the timing of physicists; direct statistical comparison with clinician contouring time is therefore inappropriate. In the cases where there were similar tasks (such as in the breast workflow) there was a slight improvement in the throughput of planning, but plan optimisation was still a large time factor.

We explain that AutoContour aided contouring and consequently expedited elements of treatment planning; the software did not interfere with treatment provision processes in the study.

Notably physicists stated that, consistent, vendor-standardised OAR delineations enhanced dose-volume histogram (DVH) review processes and made OAR tolerance tests more reproducible in plan optimisation; that is, standardized contours made inter-observer variability in OAR volumes lower and, thus, simplified comparative DVH evaluation in plan development. The practical utility of this, more reproducible OAR volumes feeding into faster and more consistent DVH checks, contributes to the reason why some physicists ranked the system highly in aiding plan development even when the amount of time saved on optimisation per se itself was small.

The quantitative time metrics are accompanied by valuable context, which is user experience and workflow impacts. Figures 5 and 6 represent workflow and user-experience outcomes. In general, the respondents said that standardised auto-generated OARs enhanced reproducibility and made the process of reviewing DVH more efficient; these advantages and the major perceived constraints are summarised in Figure 7. Figure 8A–C demonstrates accuracy and verification items, which reflect a pragmatic attitude among the staff: although a significant number of people rated the tool as quite accurate, the majority of clinicians still conduct a manual review and editing to accept the tool.

Themes that were common include technical and operational barriers. The main practical challenges reported were license availability, concurrent-user restrictions and network/TPS integration concerns (Figure 7). In terms of implementation we have found that centres are better planned to license to a peak number of simultaneous users rather than the number of users in general and that network connectivity and TPS should be up to the task to prevent DICOM import and licence-server bottlenecks. Practical steps which enhanced our deployment were a specific clinical super-user assigned to each shift, vendor update windows negotiated not to interrupt clinic hours and benchmarking of TPS I/O performance pre-deployment.

Clinical acceptance is still based on trust and accuracy. The time-savings were valued by the respondents but local clinical validation and regular QA were necessary (Figure 8A–C). The use of standardised OAR contours minimised inter-observer variability and made DVH comparisons in the development of plans easier; nonetheless, the final decision about contour acceptance and quality of the plan remained with the clinicians. This human-in-the-loop model in which AI supplements and does not substitute expert judgement is a more accurate reflection of the current safe practice.

It is worth noting the limitations of this study. It is a single-centre assessment with a small sample size and no a-priori statistical power analysis; therefore, the inferential statistics are exploratory and are to be viewed with caution. There was no formal saturation testing done and the open-ended responses were summarised in a narrative form; consequently, qualitative results are descriptive and not comprehensive. The AutoContour license constrained co-occurrence at the time of observation and possibly reduced the effect of time-savings of measurements at peak times. Lastly, timings are a combination of system time and user-reported start/stop time and may contain reporting bias even when cross-verification is attempted.

Overall, AI auto-contouring induced significant time savings in clinician contouring time at our centre and enhancing the reproducibility of OAR delineations, which had downstream effects on the planning workflow in the selected settings. The combination of respondent occupation and experience (Figure 1 and Figure 2), tumour-site specific time comparisons (Figure 3a–d), physicist planning data (Figure 4), user-experience and workflow (Figure 5 and Figure 6), benefits/limitations (Figure 7), accuracy/verification (Figure 8A-C), and paired inferential plots (Figure 9a4) give a consistent picture: AI can bring a significant efficiency enhancement but needs proper licences, strong IT infrastructure, local validation and further clinician supervision. To establish whether these time-savings can be converted to better throughput without affecting plan quality or patient safety, future multi-centre, patient-level studies, which combine dosimetric comparisons and outcome measures are needed.^13–15

Conclusion

The use of AI in radiation therapy planning is a promising tool to improve patient throughput by time-saving measures. It however, requires detailed oversight to ensure patient safety. We firmly believe that with validation, clinical implementation of AI-tools can be accomplished. Clinical trials with rigorous evaluation are needed to verify the comprehensive performance of these tools. To ensure the robustness in the AI tools, these clinical trials must have diverse enrollment to train the AI on non-standard organ morphology, prioritize personnel training on using the tools and have a focus on collaboration and constructive discussion of limitations and improvements.