In this work, we presented a deep learning-based workflow for automated pelvimetry in pre-operative MRI volumes. We showed that crucial landmarks can be localized with a mean error of 5.6 mm, which is comparable to the human inter-observer variability (3.7 ± 8.4 mm). In addition, these landmarks led to an accurate measurement of the clinically relevant pelvic dimensions. Therefore, the study results demonstrated that MRI-based metrics can be extracted automatically for difficulty assessment of the TME.

To our knowledge, this is the first study to automate pelvimetry in MRI volumes using deep learning. Due to the earlier absence of this automation, measuring these pelvic dimensions was not part of the clinical workflow. However, the dimensions provide insights into surgical access to the rectum and the associated challenges (e.g., disturbance of maneuver). Therefore, the proposed workflow can directly support the current practice as the given dimensions will enable physicians to make a more patient-specific assessment.

Several studies have explored the potential associations between pelvimetry and clinical outcomes in rectal cancer surgery, emphasizing the relevance of pelvic dimensions in surgical planning and prognosis. In previous studies, smaller pelvic dimensions, particularly the interspinous (IS) distance, intertubercle (IT) distance, and the pelvic inlet, have been associated with increased surgical difficulty and compromised outcomes. The retrospective analysis of the European MRI and rectal cancer surgery (EuMaRCS) study by d’Angelis et al. reported that a smaller IS was associated with higher conversion rates to open surgery [18]. Furthermore, Boyle et al. and Baik et al. reported that a smaller IS was associated with a positive circumferential margin [19, 20]. Similarly, reduced IT distances emerged as predictors of longer laparoscopic dissection times and increased post-operative complications in studies by Kim et al. and de’Angelis et al., highlighting its potential as an operative planning metric [12, 18]. Escal et al. demonstrated that an IT > 10.1 cm was associated with increased surgical difficulty according to their composite score [5]. Moreover, the pelvic inlet’s anterior–posterior dimension was related to critical surgical metrics, including CRM involvement, TME completeness, and intraoperative blood loss [8, 19, 20]. However, these compelling findings also reflect variability in the definitions of surgical difficulty and the pelvic measurement methods, which may influence the generalizability and interpretation of results.

Building on these findings, our study underscores the potential of automated pelvimetry on MRI as a standardized and efficient modality for pre-operative assessment of pelvic dimensions. This approach has implications for predicting surgical complexity and informing post-operative outcomes. By integrating such predictive measurements into pre-operative planning, clinicians could stratify patients based on risk to optimize surgical and perioperative strategies. In addition, it may potentially improve outcomes by tailoring interventions to individual anatomic challenges. However, our method requires further validation, and the associations with specific surgical outcomes need to be tested in prospective clinical studies to confirm its reliability in clinical practice. Also, it is essential to acknowledge that surgical complexity is inherently multifactorial, influenced not only by pelvic dimensions but also by patient- and tumor-specific factors, such as body mass index (BMI) and tumor height [21, 22]. Therefore, in addition to pelvic dimensions, a pre-operative assessment tool should also consider patient characteristics to make a comprehensive assessment. Further study will focus on applying machine learning to integrate both anatomic and clinical considerations. Such advancements can potentially support surgical planning, enhancing precision and patient-centered care.

Notably, this study found that the performance of the automated extraction can be influenced by the imaging quality. For example, the LoA and R2 values with and without outliers (Table 2) indicated that the overall performance was affected by outliers. Visual inspection of the outliers showed that most outliers in the measurement of the pelvic inlet had an MRI acquisition presenting an MRI banding artifact. As shown in Fig. 4, the artifact partially covers the os pubis and is associated with a failure of the automated landmark detection. This indicated that the model cannot make an accurate pelvic dimension determination for every MRI acquisition. In exceptional cases, such as shown in Fig. 4, the resulting erroneous dimension measurement could influence the physician’s judgment. Therefore, clinical application of the method would require a visual check or an automated recognition of outliers. A method for recognizing outliers could be developed based on the uncertainty of the model tied to the predicted landmark location or the probability of the pelvic dimension. In case of low certainty within the landmark detection or an improbable measured dimension, this can be communicated to the practitioner or a manual measurement can be chosen.

A An example visualization in which the banding artifact influences the depiction of the os pubis. B Summation of the predicted heatmaps, showing the inability of the model to localize the five landmarks

Future research should also study the broader application of automated pelvimetry. First, an external validation can be performed on a prospective dataset. This would validate the performance of the methodology at other institutes. Secondly, the presented methodology should be studied on the automated measurement of other pelvic dimensions. For example, the methodology presented can be used in transverse MRI volumes. This would enable the automated measurement of the intraspinal and intratubular distances, which have also been described as predictive for TME outcomes [6]. Thirdly, the methodology can also be used for automated pelvimetry in CT acquisitions. A comparable performance for this modality would increase the applicability of the presented methodology. Finally, automated pelvimetry can be applied to detect landmarks relevant to other clinical fields. For instance, certain pelvic dimensions have been described in the literature as predictors for the radicality of the prostatectomy [18].

In conclusion, we showed that pelvimetry in MRI volumes can be automated using deep learning. The proposed workflow can extract pelvic dimensions with high accuracy, facilitating a more patient-specific assessment of the surgical difficulty. Future research will focus on quantifying the predictive value of the pelvic dimensions to create a pre-operative assessment tool for post-operative complications.