Comment on “ <scp>SegORG</scp> : Report Generation of Oral Potentially Malignant Disorders Image Based on Lesion Segmentation‐Enhanced <scp>LLM</scp> ”

We read with great interest the recent article by Zhang et al. (2025) describing SegORG, a novel model for automated report generation of oral potentially malignant disorders (OPMDs). The authors have developed an innovative approach combining lesion segmentation with large language models, achieving impressive performance metrics. We would like to offer some constructive suggestions that may further enhance the clinical applicability of this promising work. The clinical diagnostic F1-score of 0.695 demonstrates encouraging performance. We suggest that future iterations might benefit from expanding the evaluation framework to include additional clinical scenarios, such as atypical presentations or borderline cases, which would provide valuable insights into the model's robustness across the full spectrum of OPMD manifestations. The expert evaluation revealed a median Likert score of 4 with substantial agreement at Fleiss' Kappa of 0.64, which is commendable. However, we note that the evaluation involved five randomly selected test samples. Expanding this evaluation to include a larger and more diverse sample set, perhaps stratified by lesion type, location, and severity, might provide additional confidence in the model's clinical acceptability across various presentations. The text augmentation strategy using DeepSeek-V3 to generate three distinct reports for each image is methodologically innovative. We wonder whether incorporating reports from clinicians with diverse training backgrounds and experience levels during the augmentation process might further enrich the linguistic diversity and better reflect real-world clinical documentation practices. Additionally, reporting inter-rater reliability metrics for the augmented training data could provide valuable transparency regarding data quality. The model's interpretability through lesion visualization is particularly noteworthy (Ma et al. 2024), as it aligns well with current best practices for explainable AI in medical applications. We suggest that future work might explore incorporating uncertainty quantification, allowing the system to flag cases where its confidence is low and human review is particularly warranted. Such a feature could help prevent automation bias (Schaffer et al. 2019) and maintain appropriate clinical oversight. Additionally, establishing clear usage guidelines that define appropriate clinical contexts for SegORG deployment, specify when human review is mandatory, and outline procedures for handling discrepancies between AI-generated and clinician assessments would facilitate safe and effective clinical implementation. The authors acknowledge the need for multicenter trials to assess generalizability, which we strongly support. We would additionally suggest prospective studies that evaluate how SegORG integrates into actual clinical workflows, examining factors such as time savings, clinician satisfaction, and impact on diagnostic accuracy when used as a clinical assistant tool. Assessment frameworks like TRIPOD-LLM (Gallifant et al. 2025) could provide valuable guidance for such validation studies. In conclusion, SegORG represents an important advancement in automated OPMD report generation with its efficient architecture of 90.4 million trainable parameters and superior natural language generation metrics. These suggestions are offered in the spirit of helping translate this excellent technical work into a robust clinical tool that can meaningfully support oral medicine practitioners. Zekai Yu: formal analysis, writing – original draft, writing – review and editing, investigation. Feiwei Qin: writing – review and editing, formal analysis. Use of Generative AI: During the preparation of this work the authors used Claude Opus 4.5 to improve the readability and language quality of the manuscript. After using this tool the authors reviewed and edited the content as needed and take full responsibility for the content of the published article. The authors have nothing to report. The authors have nothing to report. The authors have nothing to report. The authors declare no conflicts of interest. The authors have nothing to report.