AI–Based Segmentation of Cemento-Osseous Dysplasias: Methodological Considerations and Clinical Perspectives

Segmentation of Cemento-Osseous Dysplasias Using an Artificial Intelligence Algorithm," recently published in the International Dental Journal.The authors should be commended for addressing an underexplored area in dental artificial intelligence, namely the automated segmentation of cemento-osseous dysplasias on CBCT images.Given the diagnostic complexity and long-term follow-up challenges associated with these lesions, the attempt to standardize radiologic assessment through deep learning represents a timely and potentially impactful contribution.The study demonstrates that a nnU-Net v2-based framework can achieve reasonable segmentation performance in mixed-stage cemento-osseous dysplasias, with Dice and Jaccard values comparable to those reported for other jaw lesion segmentation tasks.From a technical perspective, the use of 3-dimensional CBCT data and voxel-level annotation is a clear strength, as it avoids the superimposition limitations inherent to 2-dimensional imaging and aligns with current trends toward volumetric AI analysis in dentomaxillofacial radiology.At the same time, the work raises several points that may stimulate further discussion and guide future research in this field.First, the exclusive inclusion of mixed-stage lesions, while methodologically understandable, narrows the clinical spectrum addressed by the model.Cemento-osseous dysplasias evolve dynamically over time, and their radiologic appearance may shift from radiolucent to radiopaque stages. 2 From a clinical standpoint, the early and late stages are often those that generate the greatest diagnostic uncertainty.Expanding AI models to encompass the full radiographic continuum of these lesions may therefore be essential if such tools are to support real-world diagnostic decision-making rather than confirm already recognizable patterns.Second, lesions were segmented as a single entity without subtype differentiation.Although segmentation per se does not require categorical labelling, cemento-osseous dysplasias comprise biologically and clinically distinct subtypes with different distributions, risks, and follow-up considerations.3 Incorporating subtype-aware annotation or post-segmentation classification could enhance the translational value of AI systems by aligning them more closely with clinical reasoning and current WHO classifications.Another point of broader relevance concerns generalizability.The dataset was derived from a single centre and a