AI-Based PET Metric Comparison in Colorectal Cancer: An Exploratory Agreement Study Between RECOMIA and SyngoVia

INTRODUCTION

Colorectal cancer (CRC) is a common malignancy prevalent globally and associated with substantial morbidity and mortality. The need for accurate staging, response assessment, and restaging of CRC tumours is of paramount importance for optimal treatment planning and patient outcomes. Radiological imaging techniques such as computed tomography (CT), magnetic resonance imaging, and positron emission tomography/CT (PET/CT) play a crucial role in these clinical tasks.^[1,2]

While standardised uptake values (SUV) on 18F-fluorodeoxyglucose (FDG) PET/CT have conventionally been utilised for tumour response assessment, they are not free of limitations such as susceptibility to image noise and alteration by patient characteristics and imaging parameters.^[3,4] An alternative and more comprehensive metric, metabolic tumour volume (MTV), offers a potential solution by considering the entire tumour burden with increased metabolism. In addition, total lesion glycolysis (TLG), a product of mean SUV (SUVmean) and MTV, provides valuable insight into tumour activity.^[5,6]

However, accurate measurement of tumour metabolic volume requires precise segmentation, a challenging task given the irregular shape of colorectal tumours and their proximity to physiological structures. Existing methods, including threshold-based techniques, suffer from limitations such as overestimation of volume due to spill-over effects and exclusion of tumours with low uptake or heterogeneity.^[4] In addition, adjacent physiological structures with high uptake interfere with the correct measurement. The necessity for a standardised, adaptable methodology for threshold determination remains unaddressed.

In recent years, the intersection of artificial intelligence (AI) and radiomics has emerged as a promising avenue for refining tumour volume assessment. The Research Consortium for Medical Image Analysis (RECOMIA) offers a cloud-based platform that leverages deep learning tools to perform organ segmentation on CT and enables precise quantitative evaluation in PET/CT images, including tumour tissue. The ability to perform both automated and manual segmentation leads to more accurate and consistent volume estimations, thereby enhancing the robustness of assessments.^[7]

This paper aims to explore the potential of AI and radiomics in addressing the challenges associated with accurate tumour volume assessment in CRC. By leveraging the capabilities of RECOMIA and its deep learning tools, we propose a novel approach to overcome the limitations of current methods and provide a more reliable and comprehensive understanding of tumour metabolic volume. This innovative strategy is likely to contribute to improving patient care and treatment decision-making in CRC.

MATERIAL AND METHODS

RESULTS

Among the 15 patients included in the study, 11 were male, and 4 were female, with a mean age of 58.2 ± 15.7 years (range: 41–79 years). This diverse cohort provided a robust basis for evaluating differences in tumour imaging metrics derived from the Syngo Via and AI-driven RECOMIA platforms. Analysis revealed significant discrepancies in SUVmean, MTV, and TLG between the two methodologies, with P values of 0.0001, 0.0003, and 0.0312, respectively. These findings highlight the superior sensitivity of AI-driven metrics in detecting and quantifying metabolic tumour characteristics. The AI methodology consistently provided broader and more nuanced evaluations, underscoring its potential for improving the precision of tumour activity assessment.

In contrast, SUVmax values showed no significant differences between the two methods, with a P = 0.2058, suggesting comparable performance in this parameter.

The observed variations in SUVmean, MTV, and TLG metrics emphasise the advanced capability of AI tools like RECOMIA to deliver a comprehensive and detailed analysis of metabolic tumour activity, positioning them as a trans formative addition to clinical and research imaging workflows.

Mean standardised uptake value

To evaluate the agreement between SUVmean values derived from Syngo Via and the AI-driven RECOMIA platform, a Bland–Altman plot was employed. The analysis revealed an average difference of 8.55 ± 6.22, with a 95% CI spanning from 5.11 to 11.99. The P value for this comparison was 0.0001, which is markedly below the conventional significance threshold of 0.05.

Close

Fig 1:

Maximum standardized uptake value Bland–Altman Scatter: Assessing Disparity in Syngo Via and artificial intelligence. SUVmax: Maximum standardized uptake value, AI: Artificial intelligence

Export to PPT

These results indicate a statistically significant difference in SUVmean values between the two methodologies. Notably, the AI-based RECOMIA platform consistently produced lower SUVmean measurements compared to Syngo Via. This suggests inherent differences in how each approach evaluates metabolic tumour activity, with AI demonstrating distinct sensitivity and precision in its assessment.

These findings underscore the potential impact of AI-driven analysis in providing alternative and potentially more nuanced evaluations of tumour characteristics. The divergence in SUVmean values also highlights the necessity for further research to elucidate the clinical implications of these differences.

Comprehensive details are available in Table 2, which outlines the comparative data points, and the variance is visually depicted in Fig 2, showcasing the Bland–Altman plot for SUVmean analysis. This robust statistical assessment emphasises the critical role of AI in refining tumour imaging methodologies.

Table 2: Mean standardised uptake value values in Syngo Via versus artificial intelligence: A Bland–Altman analysis with P-value significance

Data point	SUVmean (Syngo via)	SUVmean (AI)	Mean difference ±SD	95% CI	p
1	10.52	6.67	8.55±6.22	5.11 to 11.99	0.0001
2	19.41	7.7
3	12.17	6.35
4	9.83	5.63
5	4.58	2.82
6	38.5	14.92
7	12.13	6.87
8	16.2	6.29
9	4.18	2.83
10	22.00	8.50
11	18.30	7.20
12	14.60	5.90
13	13.80	6.00
14	5.50	3.10
15	28.70	11.40

SUVmean: Mean standardised uptake value, AI: Artificial intelligence, SD: Standard deviation, CI: Confidence interval, p < 0.05 is considered statistically significant

Close

Fig 2:

Mean standardized uptake value Bland–Altman Scatter: Assessing Disparity in Syngo Via and artificial intelligence. SUVmean: Mean standardized uptake value, AI: Artificial intelligence.

Export to PPT

Metabolic tumour volume

MTV values obtained from both Syngo Via and AI methods were evaluated using a Bland–Altman plot to visually assess their congruence. The computed average difference was −54.97 ± 44.82, with a 95% CI ranging from −79.79 to −30.15. The P = 0.0001 is well below the standard significance threshold of 0.05, indicating that the variance in MTV between the two methodologies is statistically significant. This suggests a substantial difference in how Syngo Via and AI assess MTV, with AI generally yielding higher volume measurements. For a comprehensive view, refer to Table 3 and its corresponding illustration in [Fig 3].

Table 3: Comparison of Syngo Via and artificial intelligence metabolic tumour volume values

Data point	Syngo via (MTV)	AI (MTV)	Mean difference ±SD	95% CI	p
1	8.1	35.9	−54.97±44.82	−79.79 to −30.15	0.0003
2	0.69	4.35
3	20.14	70.17
4	31.32	121.79
5	3.63	74.32
6	38.50	228.7
7	25.31	81.3
8	5.81	41.61
9	7.51	19.24
10	15.2	50.1
11	8.4	30.5
12	22.9	90.8
13	10.6	45.9
14	30.4	110.2
15	12.3	60.5

MTV: Metabolic tumour volume, AI: Artificial intelligence, SD: Standard deviation, CI: Confidence interval, p < 0.05 is considered statistically significant

Close

Fig 3:

Metabolic tumour volume Bland–Altman Scatter: Assessing Disparity in Syngo Via and artificial intelligence. MTV: Metabolic tumour volume, AI: Artificial intelligence

Export to PPT

Total lesion glycolysis

TLG measurements obtained from both Syngo Via and AI techniques were assessed using a Bland–Altman plot to visually compare their alignment. The mean discrepancy was recorded at-287.78 ± 465.41, with a 95% CI ranging from −545.51 to −30.04. The P = 0.0312 is below the conventional significance threshold of 0.05, indicating that the disparity in TLG measurements between the two methods is statistically significant. This suggests that AI tends to produce higher TLG values than Syngo Via, reflecting a notable difference in how the two methods assess tumour lesion glycolysis. For a detailed analysis, please refer to Table 4 and the corresponding visual representation in [Fig 4].

Table 4: Comparison of Syngo Via and artificial intelligence total lesion glycolysis values

Data point	Syngo via (TLG)	AI (TLG)	Mean difference ±SD	95% CI	p
1	85.21	35.9	−287.78 ±465.41	−545.51 to −30.04	0.0312
2	13.39	33.50
3	245.10	445.58
4	307.88	685.68
5	16.63	209.58
6	1482.25	3412.20
7	307.01	558.53
8	94.12	261.73
9	31.39	54.45
10	334.40	425.85
11	153.72	219.60
12	334.34	535.72
13	146.28	275.40
14	167.20	341.62
15	353.01	689.70

TLG: Total lesion glycolysis, AI: Artificial intelligence, SD: Standard deviation, CI: Confidence interval, p < 0.05 is considered statistically significant

Close

Fig 4:

Total lesion glycolysis Bland–Altman Scatter: Assessing Disparity in Syngo Via and artificial intelligence. TLG: Total lesion glycolysis, AI: Artificial intelligence

Export to PPT

DISCUSSION

This study examined how SUVmax, SUVmean, MTV, and TLG differ between an AI-based segmentation tool (RECOMIA) and conventional semi-automated software (Syngo Via) in patients with CRC. RECOMIA consistently produced lower values for SUVmean and higher values for MTV and TLG, whereas SUVmax showed comparable results between the two platforms. These findings highlight that the segmentation technique itself has a substantial impact on quantitative PET parameters.

It is important to note that higher values generated by AI-based segmentation do not necessarily represent more accurate or clinically meaningful assessments. Without a ground truth, such as histopathologic tumour volume or outcome-based validation, these differences must be interpreted with caution. Variations may arise from the way each method approaches tumour boundary definition, particularly in regions with complex anatomy or adjacent physiological uptake.

The higher MTV and TLG observed with RECOMIA likely reflect intrinsic characteristics of deep-learning-based segmentation, including the inclusion of low-uptake peripheral voxels or nearby physiologic activity. Manual editing was intentionally avoided in this study to eliminate inter-operator subjectivity and to allow an unbiased evaluation of the algorithm’s native performance. While this approach improves consistency, it may also introduce uncertainty in anatomically challenging regions. Therefore, AI-derived contours should be interpreted cautiously until validated against histopathology or outcome-based performance metrics.

Threshold-based segmentation methods, such as the 40% SUVmax approach used in Syngo Via, have well-known limitations, including spillover effects, sensitivity to tumour heterogeneity, and difficulty in delineating lesions adjacent to physiologic activity. AI-based segmentation offers potential advantages in terms of reproducibility and standardisation, but these technical improvements do not automatically translate into clinical benefit.

Although a NEMA IQ phantom comparison would provide an ideal standardised reference, the current RECOMIA deep-learning model is trained exclusively on patient datasets and does not yet support validated phantom-based segmentation workflows. For this reason, phantom analysis was not feasible within the present study. Future research incorporating NEMA IQ phantom validation is recommended once the AI platform integrates dedicated phantom segmentation capabilities.

Our results also underscore the need for larger studies to understand whether AI-derived PET metrics ultimately improve prognostic accuracy or influence treatment planning. Given the retrospective nature of this work and the relatively small sample size, caution is necessary when generalising these findings. Prospective validation, preferably with outcome-correlated data, remains essential before integrating AI-based segmentation into routine clinical workflows.

In summary, while AI tools like RECOMIA show promise for automating and standardizing tumor volume assessment, their clinical utility requires further evidence. Until robust validation is achieved, AI segmentation should be regarded as an emerging technology rather than a replacement for established clinical methods.

CONCLUSION

This study compares PET/CT-derived metrics in CRC using AI (RECOMIA) and regular software (Syngo Via) in a way that has not been done before. The method utilising AI consistently provided lower values for SUVmean and higher values for MTV and TLG. SUVmax, on the other hand, remained comparable across platforms. These results demonstrate how the method of identifying and outlining tumour areas can influence PET parameter outcomes. They also show how important it is to standardise quantitative imaging.

However, without a widely accepted standard reference for comparison or validation against clinical outcomes, the observed differences should be considered variations in measurements rather than evidence of improved accuracy or clinical utility. AI has the potential to automate and standardise tumour measurement, but further studies with robust validation are necessary to determine if these metrics genuinely enhance diagnostic performance or significantly influence patient care.

Until these metrics are thoroughly tested and confirmed, AI should be used cautiously in routine clinical practice, considering both its capabilities and current restrictions.