Letter to the Editor: why preoperative STAS prediction requires stronger pathological and imaging validation
Year: 2026
Type: letter
Abstract: HIGHLIGHTS The reliability of STAS as a pathological reference standard remains insufficient for model-driven surgical decisions. Cohort–practice mismatch limits the clinical applicability of current STAS prediction models in real operative scenarios. Complex radiomics models require benchmarking against transparent clinical scores to justify added surgical value. Lack of robust imaging harmonization raises concerns about the generalizability of multicenter radiomics performance. Preoperative STAS prediction demands stronger methodological validation before influencing resection strategy. To the Editor, We read with interest the study by Zhu et al[1], who developed a radiomics-based machine learning model to predict spread through air spaces (STAS) preoperatively, an area of increasing relevance as resection planning for early-stage lung adenocarcinoma becomes more tightly coupled to imaging-derived risk stratification. While the authors advance an important objective, several methodological and translational issues warrant careful consideration before such a model can meaningfully guide operative decision making. This study complies with the TITAN 2025 framework for transparent and ethical reporting of AI-assisted research[2]. A central concern is the reliability of the pathological reference standard. STAS has limited inter-observer reproducibility and is highly vulnerable to artefacts introduced during tissue handling, sectioning and alveolar collapse, generating label noise that constrains the achievable performance of any predictive model[3]. Without reporting pathologist concordance or detailing safeguards used to distinguish true STAS from processing mimickers, it is unclear whether the algorithm is learning generalizable biology or center-specific diagnostic variation. For a tool that could prompt surgeons to adopt more extensive resection, unquantified diagnostic instability carries direct implications for patient safety, postoperative lung function and long-term oncologic outcomes. Equally important is the divergence between the development cohort and the population in whom STAS stratification should influence treatment. More than ninety per cent of included patients had T1, node-negative tumors and were strong candidates for anatomic resection. Yet the clinical dilemma between sublobar and lobar procedures typically arises in patients with impaired reserve, borderline resectability or compromised imaging quality[4]. A model trained almost exclusively in a highly curated early-stage cohort risks overstating its relevance in routine multidisciplinary practice, where physiological heterogeneity and tumor complexity are greater and the consequences of under- or over-resection are more profound. The incremental value of the radiomics–habitat architecture also requires clearer justification. The model was not benchmarked against a transparent clinical-imaging score constructed from established predictors such as solid component proportion, lobulation, margin clarity, and serum tumor markers. Without evidence of meaningful gains in discrimination, calibration, reclassification or clinical utility, the rationale for adopting a high-dimensional algorithm over an interpretable logistic model remains uncertain[5]. Given the irreversible consequences of surgical decisions, opacity without demonstrable benefit raises practical and ethical concerns, particularly regarding interpretability, accountability and alignment with multidisciplinary decision making. Finally, generalizability is constrained by the lack of rigorous harmonization across scanners and acquisition protocols. Radiomic features are acutely sensitive to reconstruction kernel, voxel geometry, and vendor-specific intensity patterns and may encode acquisition settings rather than tumor biology if uncorrected. In the absence of phantom calibration, kernel normalization or batch-effect adjustments such as ComBat, the reported multicenter performance may reflect protocol congruence more than true robustness[6]. A model intended for operative planning must demonstrate resilience to the variability inherent in real-world imaging ecosystems. Taken together, these limitations do not diminish the value of pursuing preoperative STAS prediction, but they indicate that the model is best regarded as hypothesis-generating rather than practice-changing. Strengthening pathological reference standards, aligning development cohorts with the patients in whom resection extent is genuinely contested, benchmarking complex algorithms against transparent clinical models and embedding robust imaging harmonization will be essential before such tools can responsibly influence surgical strategy. Integrating radiomics-based STAS stratification into risk-adapted lung cancer surgery ultimately requires not only technical innovation but also methodological rigor, interpretability and clear lines of clinical responsibility.
Source: International Journal of Surgery
Language: English
Subfield: Radiology, Nuclear Medicine and Imaging
Field: Medicine
Domain: Health Sciences
Open Access status: hybrid
Estimated APC paid: $3,450
OpenAlex ID: W7124243001