Letter to the Editor: Deep learning-driven multi-hierarchical granularity integration for surgical scene understanding: experimental study

Dear Editor,

We commend the recent paper by Chu et al^[1], which proposes a multi-granularity deep learning framework for surgical scene understanding during laparoscopic radical nephrectomy. This work integrates surgical stage, step, instrument, and action recognition with a hemorrhage risk prediction module, marking a significant technical breakthrough in this field. To improve these findings’ interpretive value and clinical applicability, we would like to make a few remarks, as outlined in the Transparency In The reporting of Artificial Intelligence – The TITAN guideline^[2].

First, the data set comprises procedures from 10 surgeons across two institutions, with no stratification by experience level^[3]. This risks confounding AI-predicted bleeding risk with operator-specific skill patterns. Junior surgeons often have prolonged operative times and minor, routine oozing – labeled “high risk” by the AI under the authors’ threshold – whereas experts may perform rapid, controlled dissections that visually appear aggressive but pose minimal real risk.

Second, the 20-mL estimated blood loss cutoff for high-risk events was based on the consensus of five urologists but is not validated against patient outcomes nor supported by established surgical literature. Such a low threshold may inaccurately classify benign episodes as clinically significant^[4].

Third, the multi-task architecture has millions of parameters but was trained on n = 41 patients, without an a priori power analysis. The authors report testing 123 model configurations without correction for multiple comparisons, elevating the likelihood that “optimal” results arose from statistical noise^[5].

In addition, drawing from the original paper’s reported fold-level mean ± SD, we note that the bleeding risk prediction module achieved modest improvements over baseline across all metrics, but with substantial variance. We generated a comparative bar chart (Fig. 1) reproducing these metrics with standard deviation error bars to illustrate fold-level variability.

Figure 1.

Baseline vs proposed model performance for intraoperative bleeding risk prediction in Chu et al, with error bars denoting fold-level standard deviation. Data reproduced exactly from the original paper.

Despite consistent directional improvements, fold-level SDs are sizeable (7–8%), and ROC–AUC gains are small (+0.039), falling below performance levels typically required for high-stakes intraoperative decision-making. Without correction for multiple hypothesis testing across >100 trials, these margins may reflect statistical fluctuation rather than robust generalization.

In summary, Chu et al’s framework provides valuable proof-of-concept for multi-task surgical artificial intelligence. However, its clinical translation – particularly regarding hemorrhage risk prediction – is constrained by several factors: unstratified surgeon variability, arbitrary label thresholds, limited data volume confined to a single institution, absence of adjusted statistical testing, and modest performance gains. Prior to deployment, prospective multicenter validation using empirically stratified datasets and clinically validated endpoints is essential.

Ethical approval

None.

Consent

None.

Sources of funding

This work received no specific grant from any funding agency, commercial, or not-for-profit sectors.

Author contributions

Z.W. contributed to the study concept and design, performed data collection, and participated in drafting the manuscript. Q.L. conducted data analysis and interpretation, and contributed to drafting and revising the manuscript for important intellectual content. W.D. supervised the study, contributed to study design, and provided critical revision of the manuscript. All authors approved the final version of the manuscript and agree to be accountable for all aspects of the work.

Conflicts of interest disclosure

The authors declare no conflict of interest.

Research registration unique identifying number (UIN)

Not applicable, because this is correspondence, providing commentary on existing literature without involving any original research, human or animal data, participants, or studies.

Guarantor

Deng Wangdong.

Provenance and peer review

Not commissioned.

Data availability statement

As no datasets were generated or analyzed in this study, data sharing is not applicable here.