A Light-Weight Practical Framework for Feces Detection and Trait Recognition.

Area of Science:

• Medical imaging
• Computer-aided diagnosis
• Digestive disease diagnostics

Background:

• Fecal trait examinations are crucial for diagnosing digestive diseases but lack automated systems due to privacy concerns and computational limitations.
• Existing object detection methods struggle with variable feces shape and location, complicating automated analysis.
• The need for efficient, low-resource diagnostic tools is high in clinical settings.

Purpose of the Study:

• To develop a practical, lightweight framework for automatic feces detection and trait recognition.
• To overcome challenges related to image acquisition, variable object properties, and computational constraints in automated fecal analysis.
• To improve the accuracy and efficiency of fecal examinations for clinical diagnosis.

Main Methods:

• A three-stage framework: illumination normalization, feces detection, and trait recognition.
• A threshold-based segmentation scheme for feces detection, independent of training and labeling.
• A lightweight, shallow convolutional neural network (CNN) for trait classification.

Main Results:

• The proposed framework achieved a satisfactory accuracy of 98.4% on a collected dataset.
• The system demonstrated low computational complexity and storage requirements.
• Illumination normalization effectively reduced accuracy degradation due to lighting variations.

Conclusions:

• The developed lightweight framework offers a practical solution for automated feces detection and trait recognition.
• This system is suitable for real-world hospital environments, particularly mobile computer-aided diagnosis devices.
• The approach addresses key challenges, paving the way for improved diagnostic capabilities in digestive health.