- Conference Article
- 10.1109/bibe66822.2025.00037
Classifying Pre-Malignant Colon Polyps Using Hybrid Deep Learning on Ex Vivo Optical Coherence Tomography Images
- Nov 06, 2025
- Christos Photiou + 3 more +3
To alleviate the workload associated with colorectal polyp screening, a leave-in-situ approach for diminutive polyps (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\leq 5 \text{mm}$</tex>) is under consideration. This strategy, however, necessitates enhanced diagnostic performance in line with the PIVI-1 criterion. This study aimed to employ Optical Coherence Tomography (OCT) to differentiate colorectal polyps as either benign (normal or hyperplastic) or malignant potential (adenoma or sessile serrated adenoma (SSA)), achieving a level of accuracy suitable for clinical colorectal cancer screening. OCT image volumes <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(1000 \times 1000 \times 1024$</tex> pixels) were acquired from polyps from 225 patients. En face and feature-enhanced en face images were created. The en face images were annotated by an expert histopathologist. The resulting datasets were used to train individual U-Net models to identify malignant potential. Following U-Net training, a two-stage fusion process was applied, aggregating the classification scores from multiple processed datasets and employing a range of traditional machine learning classifiers to achieve the final decision. The area under the curve (AUC) for the detection of malignant potential of all polyps was 0.90. In addition, the classification achieved an accuracy of 90.3%, sensitivity of 96.9%, and specificity of 81.5%. The results of this study confirm the potential of OCT imaging of colorectal polyps as a viable adjunct to colonoscopy that could enable leave-in-situ management strategies.
Read more
22