Computer-Aided Endoscopic Diagnosis Without Human-Specific Labeling.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Supervised machine learning for endoscopic diagnosis requires tedious pixelwise labeled data.
• Collecting precisely labeled endoscopic images is time-consuming and labor-intensive.
• Endoscopic medical reports offer readily available weakly labeled data.

Purpose of the Study:

• To develop a computer-aided endoscopic diagnosis system that utilizes weakly labeled data from diagnostic text.
• To design a system that does not require human-specific pixelwise labeling.
• To create a novel system suitable for clinical settings.

Main Methods:

• Representing endoscopic images and reports as bags and instances, using a global bag-of-words model.
• Employing a feature mapping scheme to identify the most suspicious lesion instance from positive bags.
• Utilizing an online metric learning method for self-updating classification with new data.

Main Results:

• Achieved an Area Under the Curve (AUC) of 0.93 on a dataset of over 12,000 images from 424 volunteers.
• Demonstrated superior performance compared to state-of-the-art methods.
• Showcased improved information mining from unseen bags through an online phase.

Conclusions:

• The developed system effectively utilizes weakly labeled endoscopic images for diagnosis.
• The novel approach outperforms existing methods by mining positive instances and employing online learning.
• Presents the first weakly labeled endoscopic image dataset and a clinically applicable diagnostic system.