A fusion model of manually extracted visual features and deep learning features for rebleeding risk stratification in peptic ulcers.

Abstract

Objectives: We propose a multi-feature fusion model based on manually extracted features and deep learning features from endoscopic images for grading rebleeding risk of peptic ulcers.

Methods: Based on the endoscopic appearance of peptic ulcers, color features were extracted to distinguish active bleeding (Forrest I) from non-bleeding ulcers (Forrest II and III). The edge and texture features were used to describe the morphology and appearance of the ulcers in different grades. By integrating deep features extracted from a deep learning network with manually extracted visual features, a multi-feature representation of endoscopic images was created to predict the risk of rebleeding of peptic ulcers.

Results: In a dataset consisting of 3573 images from 708 patients with Forrest classification, the proposed multi-feature fusion model achieved an accuracy of 74.94% in the 6-level rebleeding risk classification task, outperforming the experienced physicians who had a classification accuracy of 59.9% (P<0.05). The F1 scores of the model for identifying Forrest Ib, IIa, and III ulcers were 90.16%, 75.44%, and 77.13%, respectively, demonstrating particularly good performance of the model for Forrest Ib ulcers. Compared with the first model for peptic ulcer rebleeding classification, the proposed model had improved F1 scores by 5.8%. In the simplified 3-level risk (high-risk, low-risk, and non-endoscopic treatment) classification task, the model achieved F1 scores of 93.74%, 81.30%, and 73.59%, respectively.

Conclusions: The proposed multi-feature fusion model integrating deep features from CNNs with manually extracted visual features effectively improves the accuracy of rebleeding risk classification for peptic ulcers, thus providing an efficient diagnostic tool for clinical assessment of rebleeding risks of peptic ulcers.