An Accurate Deep Learning-Based System for Automatic Pill Identification: Model Development and Validation.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Pharmacology

Background:

• Medication errors are a significant cause of medical errors.
• Patients often discard medication containers, hindering accurate identification.
• This leads to potential misuse and adverse drug events.

Purpose of the Study:

• To develop a deep learning system for accurate prescription pill identification.
• To reduce medication errors by enabling real-time pill recognition.
• To assist patients and healthcare professionals in medication management.

Main Methods:

• A deep learning system comprising pill recognition and retrieval steps was developed.
• Image classification and text detection models were used for feature and imprint recognition.
• A novel language model and coordinate encoding technique were introduced for imprint correction.

Main Results:

• The system achieved high accuracy in identifying pills from South Korean (85.6%) and US (74.5%) databases.
• An accuracy of 78% was achieved with consumer-submitted images, even with limited training data.
• An ablation study confirmed the significant contribution of the language model to identification accuracy.

Conclusions:

• The proposed AI system demonstrates high precision in real-time pill identification, offering a solution to reduce medical errors.
• This technology can minimize medication misuse by patients and free up healthcare staff for more complex tasks.
• The system's ability to identify new pills without retraining highlights its practical utility in diverse healthcare settings.