Letter to the Editor Regarding the Article “Comparison of Transfer Learning Models in Pelvic Tilt and Rotation Measurement in Pediatric Anteroposterior Pelvic Radiographs”

Dear Editor,

I am writing to express concerns about a study published in your esteemed journal titled “Comparison of Transfer Learning Models in Pelvic Tilt and Rotation Measurement in Pediatric Anteroposterior Pelvic Radiographs” by Professor Huang et al. [1]. As a medical imaging researcher at the University of Sydney, I have been closely examining the methodologies employed in the field of AI-assisted radiographic analysis. My concerns pertain to the experimental design and data validation methods used in the aforementioned study.

The study in question employed CT images from 30 patients to create 2430 simulated X-ray images, which were then allocated in a 2:1:1 ratio across training, validation, and test datasets for AI model training. However, each of these datasets appears to have incorporated skeletal features from the same 30 patients. This methodology raises reliability concerns, as having all three datasets—training, validation, and test—include complete anatomical information from the same 30 patients significantly heightens the risk of overfitting the AI model to these specific cases. As a result, the model’s ability to generalize effectively to new patient data may be significantly compromised.

This issue surfaced from a systematic review titled “Evaluating Pelvic Tilt Using Pelvic Antero-Posterior Projection Images,” which demonstrated a strong correlation in predicting changes in pelvic tilt across various postures from antero-posterior projection images of a single patient [2]. This scenario closely resembles that of the study in question, yielding similar outcomes [2, 3]. Importantly, these findings pertain exclusively to relative pelvic tilt: the variation in pelvic tilt angles when the angle from one antero-posterior radiograph is known, thus enabling reliable calculations of other pelvic tilt angles from the same patient using most parameters from antero-posterior radiographs. On the other hand, an analysis of 27 peer-reviewed studies examining the correlation between absolute pelvic sagittal tilt and tilt inferred from antero-posterior radiographs revealed different outcomes. The most favorable study reported a correlation coefficient of 0.98 (n = 2) and the error of ± 2 standard deviation as 7.7° (n = 20) [2, 4]. However, this analysis also highlighted a substantial maximum error range, approximately between −10° to 12°, and a relatively small sample size [4]. Given these findings, it seems improbable that the AI model developed in the study in question could rival these established studies in accurately inferring absolute pelvic tilt angles.

Thus, the study in question incorrectly attributed these findings to absolute pelvic tilt, reporting a mean absolute error of 0.52° between AI and manually measured pelvic tilt angles [1]. This conclusion was drawn under the assumption that the AI accurately inferred an absolute pelvic tilt angle, closely matching the pre-defined absolute pelvic orientation. Yet, this overlooks a critical aspect: clinical studies in this field indicate that while the known pelvic orientation of a patient on an antero-posterior pelvic image can help calculate the pelvic orientation from other antero-posterior images of the same patient, this methodology cannot be reliably extended to different patients [2]. Consequently, the study in question overestimated the AI model’s capability to accurately predict absolute pelvic tilt values across a diverse patient population. Therefore, the results might not be replicable on patients beyond the 30 included in the training dataset, highlighting a potential limitation in the study’s generalizability and inaccurate conclusion.

I propose that the research team conduct further experiments using data from new patients to evaluate the current model’s reliability and its ability of generalization. Such research would enhance our understanding of the model’s applicability beyond the specific patient cohort used in its training dataset. Alternatively, a revision of the study’s conclusions could potentially be warranted. It might be beneficial to adjust the AI methodology to effectively utilize anatomical information from one or a few antero-posterior pelvic radiographs with known pelvic tilt angles. This adjustment could enable the fine-tuning of the AI model to predict the absolute pelvic tilt angle for the same patient from ongoing antero-posterior pelvic radiographs, potentially reducing the need for continuous sagittal pelvic radiographs for that patient.

I hope this letter will foster constructive dialog and contribute to the advancement of research in this field. Any comments and clarifications from the authors would be most welcome.

Sincerely