Endo-DET: A Domain-Specific Detection Framework for Multi-Class Endoscopic Disease Detection | JoVE Visualize

Area of Science:

Medical Imaging
Computer Vision
Gastroenterology

Background:

Gastrointestinal cancers represent a significant global health burden, necessitating effective early detection methods.
Endoscopic disease detection is crucial for reducing mortality, but faces challenges like feature redundancy, illumination inconsistency, and scale variability.
Existing methods struggle with the complex visual data encountered during endoscopy.

Purpose of the Study:

To introduce Endo-DET, a novel domain-specific detection framework for multi-class endoscopic disease detection.
To address key challenges including feature redundancy, illumination variations, and scale differences in endoscopic images.
To provide a robust and efficient solution for computer-aided diagnosis in gastroenterology.

Main Methods:

Developed the Adaptive Lesion-Discriminative Filtering (ALDF) module using sparse simplex projection to focus on lesion-specific features.
Implemented the Global-Local Illumination Modulation Neck (GLIM-Neck) for illumination-aware multi-scale feature fusion across different endoscopic modalities.
Designed the Lesion-aware Unified Calibration and Illumination-robust Discrimination (LUCID) module for integrating texture and semantic information while mitigating artifacts.

Main Results:

Endo-DET demonstrated significant improvements in mean Average Precision (mAP) and Recall across multiple benchmark datasets (EDD2020, Kvasir-SEG, PolypGen2021, CVC-ClinicDB).
Achieved substantial gains in mAP50-95 (up to 10.8 pp) and mAP75 (up to 10.3 pp) compared to the DEIM baseline.
Showcased real-time processing capability at 330 FPS with TensorRT FP16 optimization, maintaining high accuracy.

Conclusions:

Endo-DET effectively addresses critical challenges in endoscopic image analysis, offering superior performance.
The framework provides a robust methodological foundation for real-time clinical computer-aided diagnosis of gastrointestinal diseases.
This advancement has the potential to improve early detection rates and patient outcomes in gastrointestinal cancer screening.