Recent molecular studies have suggested a biological link between hyperplastic polyps (HPs) and sessile serrated lesions (SSLs), as early serrated features and BRAFV600E-mutations are frequently found in specific HP subtypes: goblet cell-rich HPs (GCHPs) and microvesicular HPs (MVHPs). This supports the hypothesis that GCHPs, MVHPs and SSLs belong to the serrated neoplasia pathway. In previous work, we demonstrated that deep learning can differentiate GCHPs from MVHPs based on their endoscopic appearance. The present study builds further on this finding by investigating whether deep learning can also distinguish GCHPs, MVHPs and SSLs.

High-quality white light and i-scan virtual chromoendoscopy videos of polyps, acquired in our center with Pentax EC38-i20c and EC34-i10 endoscopes, were retrospectively collected, together with their histopathological analysis. Expert pathologists reviewed all cases and reclassified HPs into GCHPs and MVHPs, resulting in a dataset of 260 polyps from 194 patients: 138 GCHPs, 54 MVHPs, and 68 SSLs. The dataset was split into training (70%), validation (10%) and test (20%) sets. Three binary classifiers (GCHP vs MVHP, GCHP vs SSL and MVHP vs SSL) and one ternary classifier (GCHP vs MVHP vs SSL) were trained identically using an EfficientNet B0 backbone with a single video frame as input. Oversampling was applied to balance underrepresented classes. When applying these models to the test cases, predictions for different video frames were aggregated using majority voting.

The test set contained 28 GCHPs, 14 MVHPs and 15 SSLs. The binary GCHP-MVHP model scored 82.7% AUC and 71.4% sensitivity for both classes. The GCHP-SSL model performed similarly with 82.1% AUC and sensitivities of 64.3% and 86.7% for GCHP and SSL respectively. The MVHP-SSL model achieved the highest performance with 87.6% AUC and sensitivities of 71.4% and 80% for MVHP and SSL respectively. In contrast, the ternary model yielded sensitivities of 25%, 64.3% and 60% for GCHP, MVHP and SSL respectively, with GCHP mostly misclassified as MVHP and SSL as GCHP.

Our findings demonstrate that deep learning can differentiate GCHPs, MVHPs and SSLs based on their endoscopic appearance using binary models. The consistently high sensitivity for SSL across the binary models suggests that SSLs have more distinct endoscopic features while differences between GCHP and MVHP are more subtle. The ternary model struggled to reliably discriminate the three classes directly, likely due to the complexity of the task and the limited dataset used. Overall, this study highlights the potential of AI to support endoscopists in recognising a broader spectrum of serrated polyp types, further enhancing diagnostic precision.