A Lightweight Hybrid Representation Learning Framework for Colorectal Lesion Classification in Colonoscopy Images

Accurate colorectal lesion classification is important for computer-aided colonoscopy, yet many existing approaches require computationally intensive architectures. This study proposed and experimentally evaluated LHRF, which integrates BoVW encoding, SAE compression, TE-based metric refinement, optional HC augmentation, and FC-ResNet classification. Two configurations were investigated: LHRF-Core (BoVW → SAE → TE → FC-ResNet) and LHRF-HC, which additionally incorporates clustering-derived features. The experiments included three study-specific binary classification tasks constructed from CVC-ClinicDB, CVC-ColonDB, and a merged Kvasir/Kvasir-SEG collection. The CVC tasks used study-assigned sequence-level pathology categories, with sequence-grouped evaluation for CVC-ClinicDB and source-grouped evaluation for CVC-ColonDB. LHRF-HC achieved the highest accuracy on CVC-ClinicDB (96.24%, AUC 0.9741, MCC 0.9248), whereas LHRF-Core achieved the best results on CVC-ColonDB (95.88%, AUC 0.9698, MCC 0.9176) and Kvasir/Kvasir-SEG (98.74%, AUC 0.9908, MCC 0.9746). Five-fold cross-validation on CVC-ClinicDB yielded an accuracy of 97.43 ± 1.29%. Under matched experimental protocols, LHRF achieved higher classification accuracy than the evaluated lightweight baselines while requiring approximately 0.45–0.51 million trainable parameters. Measured post-decoding CPU throughput ranged from 351.67 to 2353.97 FPS, excluding image decoding and disk input/output. HC improved accuracy on CVC-ClinicDB but slightly reduced it on the other datasets, indicating a dataset-dependent contribution. The study also specified a gradient-based keypoint-attribution procedure, although its empirical faithfulness and clinical interpretability require further evaluation. Overall, the experimental findings highlight the potential of compact hybrid representation learning for colorectal lesion classification. Further evaluation using independently verified pathology annotations, patient-independent partitions, source-matched datasets, and external clinical collections would strengthen the assessment of robustness and clinical generalizability.