Endoscopy-based IBD identification by a quantized deep learning pipeline.

Area of Science:

• Medical Imaging Analysis
• Machine Learning in Healthcare
• Computational Pathology

Background:

• Differentiating inflammatory bowel diseases (IBD) from healthy controls using endoscopic imaging presents a significant challenge for machine learning models.
• This study utilizes this classification task as a benchmark to evaluate a novel deep learning pipeline.

Purpose of the Study:

• To develop and assess a deep learning classification pipeline for IBD detection from endoscopic images.
• To enhance model characteristics including reproducibility, interpretability, reduced computational workload, bias-free modeling, and robust image preprocessing.

Main Methods:

• An automated preprocessing procedure was developed to eliminate artifacts from clinical endoscopic data.
• A per-patient model was employed, mimicking clinical decision-making processes using aggregated image data and resampling techniques.
• Quantization was implemented to reduce model complexity and computational cost, enabling deployment on low-power devices.

Main Results:

• The pipeline achieved a Matthews Correlation Coefficient of 0.9 on a private dataset comprising 758 IBD patients and 601 healthy controls.
• Quantization resulted in a negligible 3% performance degradation, demonstrating its feasibility for real-time clinical support in resource-limited settings.
• The system provides explanations by highlighting the specific images used for prediction, aiding clinician verification.

Conclusions:

• A comprehensive preprocessing pipeline is crucial for artifact removal and effective clinical data analysis.
• Emulating clinician decision-making processes enhances model explainability and trust in healthcare applications.
• Quantization is a viable strategy for reducing computational demands in medical AI, paving the way for real-time diagnostic support tools during endoscopy.