Colorectal laterally spreading tumors (LST) are a distinct form of non-polypoid colorectal neoplasia that extend laterally along the mucosal surface and often exceed 20 mm in diameter. Despite advances in optical imaging technologies, diagnostic accuracy for large LSTs varies widely among classification systems, emphasizing the need to assess combined endoscopic approaches and recurrence prediction models to improve risk stratification and treatment planning.

Aim. To compare the diagnostic accuracy of combinations of combined endoscopic classifications and recurrence prediction models to identify the most effective approach for granular and non-granular subtypes of laterally spreading tumors.

Materials and methods. A single-center mixed retrospective-prospective study was conducted at the Medical Educational and Scientific Center “University Clinic” (Zaporizhzhia), including 110 patients with LSTs ≥20 mm (2015–2024). Granular (LST-G) and non-granular (LST-NG) lesions were assessed using JNET, Kudo, Modified Sano, and Hiroshima classifications. Six combinations of endoscopic classification systems were tested in parallel to determine diagnostic metrics. The histological evaluation of the resected neoplasia served as the reference standard. Resection techniques included endoscopic mucosal resection (EMR), piecemeal EMR, endoscopic submucosal dissection (ESD), and hybrid ESD. Recurrence was assessed at 6 months, with its prediction evaluated using the SMSA, SERT, and Baylor College of Medicine (BCM) models.

Results. The JNET + Hiroshima combination showed the highest diagnostic performance (LST-G: 81.82 % (95 % CI, 67.29–91.81 %) sensitivity, 90.91 % (95 % CI, 70.84–98.88 %) specificity, 84.29 % (95 % CI, 72.76–92.30 %) diagnostic accuracy; LST-NG: 86.67 % (95 % CI, 59.54–98.34) sensitivity, 100 % (95 % CI, 86.28–100.00) specificity, 95.00 % (95 % CI, 83.08–99.39 % diagnostic accuracy). JNET + Kudo served as a strong alternative. Progressive histological changes and recurrence were significantly more common among LST-G (68.6 %) than LST-NG (37.5 %). Recurrences were observed only in the LST-G group (8/70). BCM score ≥1 demonstrated the highest predictive ability for recurrence (AUC: 0.78), outperforming SMSA and SERT models which demonstrated poor discrimination (AUC: 0.31–0.37).

Conclusions. The combination of JNET + Hiroshima or JNET + Kudo classifications optimizes histologic prediction in both granular and non-granular large LSTs. LST-Gs demonstrate a higher risk for advanced histology and recurrence. The BCM model is preferable for recurrence prediction in large granular LSTs.