Semi-Supervised Segmentation Framework for Gastrointestinal Lesion Diagnosis in Endoscopic Images

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Machine Learning

Background:

• Accurate gastrointestinal (GI) lesion segmentation is critical for diagnosing digestive tract diseases.
• Automatic segmentation in endoscopic images aids physicians and improves patient outcomes.
• Limited pixel-wise annotations and the need for large labeled datasets hinder deep learning model development for computer-aided diagnosis (CAD) systems.

Purpose of the Study:

• To propose a semi-supervised segmentation framework using generative adversarial learning for GI lesion diagnosis.
• To address the challenge of limited annotations by leveraging both limited annotated and large unlabeled endoscopic image datasets.
• To enhance the generalizability and accuracy of CAD systems for GI lesion detection.

Main Methods:

• Developed a generative adversarial learning-based semi-supervised segmentation framework.
• Integrated limited annotated data with large unlabeled datasets for network training.
• Conducted extensive testing on a dataset of 4880 endoscopic images.

Main Results:

• Achieved superior performance compared to existing methods on challenging multi-sited datasets.
• Reported high accuracy metrics: Dice similarity coefficient (89.42 ± 3.92), Intersection over union (80.04 ± 5.75), Precision (91.72 ± 4.05), Recall (90.11 ± 5.64), and Hausdorff distance (23.28 ± 14.36).
• Validated the effectiveness of the proposed framework in improving lesion segmentation.

Conclusions:

• The semi-supervised method effectively utilizes unlabeled endoscopic images to boost lesion segmentation accuracy.
• Experimental results demonstrate the method's potential and superiority over current related works.
• The proposed CAD system shows promise in minimizing diagnostic errors in GI endoscopy.