Related Experiment Video
Updated: Apr 11, 2026
12:18
A Machine Learning Approach to Design an Efficient Selective Screening of Mild Cognitive Impairment
Published on: January 11, 2020
8.2K
Code-Based Versus AutoML Methods for Pill Recognition in Clinical Settings: Comparative Performance Study.
Amir Reza Ashraf1, Richárd Rádli2, Zsolt Vörösházi2
1Department of Pharmaceutics, Faculty of Pharmacy, University of Pécs, Rókus utca 4, Pécs, H-7624, Hungary, +36 72503650 ext 28841.
JMIR Medical Informatics
|April 10, 2026
Summary
Comparing automated machine learning (AutoML) and code-based pill recognition, this study found no single platform excelled across all clinical settings. Careful selection based on specific needs and rigorous validation are crucial for safe patient care.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Health Informatics
Background:
- Medication identification errors pose significant risks to patient safety in clinical settings.
- Automated pill recognition using computer vision offers a potential solution to mitigate these errors.
- Limited comparative studies exist for code-based versus no-code (AutoML) approaches in pill recognition.
Purpose of the Study:
- To evaluate and compare code-based (YOLOv11) and AutoML platforms for pill recognition.
- Assess performance, cost, usability, and deployment feasibility across diverse datasets, including clinical images.
- Determine the impact of training data size on model performance.
Main Methods:
- Trained models using Ultralytics YOLOv11 and three cloud-based AutoML platforms (Amazon Rekognition, Google Vertex AI, Microsoft Azure Custom Vision).
- Utilized five training subset sizes (1,230 to 26,880 images) from 30 common medications.
- Evaluated models on six datasets, including clinical, verification, and laboratory images, measuring accuracy, precision, recall, and mAP.
Main Results:
- No platform dominated all environments; accuracy varied significantly.
- On verification data, Google Vertex AI achieved up to 91.60% accuracy, while YOLOv11 showed consistent improvement with data size (up to 80.83%).
- Clinical dataset accuracy fluctuated (20.62%-90%); costs and training times varied, with YOLOv11 being open-source and others ranging from $5.43 to $69.30.
Conclusions:
- YOLOv11 offers flexibility and low cost but requires technical expertise.
- AutoML platforms provide high performance at higher costs with less control and potential unpredictability.
- Platform selection must align with specific requirements, budget, and resources, followed by rigorous real-world validation for clinical safety.
