Comment on “Deep learning habitat radiomics based on ultrasound for predicting preoperative locally progression and postoperative recurrence risk of thyroid cancer: a multicenter study”

To the Editor, Zhang et al report a large, multicenter study integrating intratumoral “habitat” radiomics and peritumoral deep learning (DLH) on ultrasound to predict locally advanced thyroid cancer (LATC) and postoperative recurrence1. In an era where artificial intelligence (AI) is rapidly entering endocrine surgery, this work is both timely and highly relevant. In preparing this commentary, we followed the TITAN 2025 recommendations for transparent and responsible reporting of surgical research that may involve AI-assisted tools2. The authors analyze 1881 retrospectively collected thyroid cancer cases from nine centers plus a 130-patient prospective cohort. LATC (extrathyroidal extension or nodal disease) is correctly identified preoperatively with impressive discrimination: the combined clinical–DLH nomogram reaches area under the receiver operating characteristic curve (AUCs) of 0.852, 0.897, and 0.906 in the internal, external, and prospective cohorts, respectively (Fig. 4, Fig. 5; Table 4)1. Importantly, DLH remains an independent predictor of recurrence-free survival, and an online Cox-based nomogram is provided for individualized recurrence risk estimation. Several aspects of this study stand out. First, it addresses a real surgical pain point: underestimation of LATC on conventional ultrasound may lead to inadequate lymphadenectomy or suboptimal initial surgery, with higher re-operation risk3,4. By combining intratumoral habitat features with a 3-mm peritumoral band, the model captures subtle patterns of capsular invasion, stromal remodeling, and lymphatic involvement that are difficult for human readers to quantify. Second, the authors go beyond AUC reporting to provide calibration, decision-curve analysis, and subgroup analyses across tumor size, age, and extrathyroidal extension/ lymph node metastasis strata (Fig. 6), supporting robustness across clinically relevant scenarios1. Nevertheless, some issues temper immediate clinical adoption. Ultrasound acquisition and segmentation were performed by experienced radiologists under controlled conditions; performance in lower-volume centers or with heterogeneous equipment is uncertain. Although external and prospective validation are strengths, all centers are within one health care system, and multifocal tumors were excluded, which may limit generalizability1. The recurrence model is based on relatively few events and requires longer follow-up. In addition, practical barriers – including the need for standardized ultrasound protocols, computational infrastructure, and integration into PACS or reporting systems – must be solved before widespread deployment, especially in resource-limited settings. For thyroid surgeons and multidisciplinary teams, the main clinical message is conceptual: ultrasound-based AI models can move LATC assessment from subjective pattern recognition toward quantitative risk estimation, potentially refining the choice between lobectomy and total thyroidectomy with compartmental dissection in borderline cases4,5. In the near term, the online nomograms are best viewed as decision-support tools to be tested alongside guideline-based assessment rather than as stand-alone arbiters. Future priorities include multi-ethnic external validation, head-to-head comparison with expert radiologists, evaluation of impact on surgical planning and re-operation rates, and prospective trials embedding DLH-based nomograms within American Thyroid Association/ National Comprehensive Cancer Network -concordant pathways. If such tools can be shown to safely downstage low-risk patients while more reliably flagging truly high-risk LATC, they may meaningfully improve both oncologic and functional outcomes in differentiated thyroid cancer.