Segmentation for classification of gastroenterology images

Area of Science:

• Medical Imaging
• Computer Vision
• Gastroenterology

Background:

• Automatic classification of gastric cancer lesions from endoscopic images is challenging.
• Accurate segmentation is crucial for reliable classification of cancerous, pre-cancerous, and normal tissues.

Purpose of the Study:

• To assess the impact of different segmentation algorithms (mean shift, normalized cuts, level-sets) on cancer classification accuracy.
• To compare classification performance using color and texture features across chromoendoscopy and narrow-band imaging modalities.

Main Methods:

• Applied mean shift, normalized cuts, and level-set segmentation algorithms to gastric tissue images.
• Utilized hue-saturation histograms and local binary patterns for feature extraction.
• Classified images from chromoendoscopy and narrow-band imaging into cancerous, pre-cancerous, and normal categories.

Main Results:

• Mean-shift segmentation demonstrated robust performance with minimal classification degradation (6%).
• Full image classification proved highly inaccurate, emphasizing the critical role of segmentation.
• Patch Index was identified as a valuable metric for assessing the classification potential of segmented regions.

Conclusions:

• Segmentation algorithms significantly influence the accuracy of automatic cancer detection in gastroenterology.
• Mean-shift offers a promising approach for lesion segmentation in endoscopic imaging.
• Further research in segmentation is essential for advancing automated diagnostic tools in gastroenterology.