Article type
Abstract
Background
The issues of drug omissions and incorrect drug identification are great important in transition of care. Pharmacists face challenges in visually identifying many unmarked and unlabeled tablets.
Objectives
This study aims to establish an image database using artificial intelligence (AI) deep learning to replace manual recognition by the human eye, especially for medications that are difficult to distinguish.
Methods
(1) Database establishment: We choose 10 pills (in supporting documents) and each medication tablet is photographed 15 images, capturing the front view and the other sides. The angles are not restricted. We also employ a ruler-backed platform, measured in centimeters to enhance the accuracy of deep learning. (2) Training and validation: We utilize EfficientNet-b5 as the training model and the model type is multi-class classification. Random-resized-crop, random-horizontal-flip, color-jitter and random-rotation are applied for data augmentations. Normalization of images is applied as well. (3) Inference: We use the Inferencer object for inference, obtains the results, and prints them in JSON format by Python. The packages include simplejson, numpy, torch and so on. All processes are coded using Python. (4) Result: The primary result of training is accuracy. In terms of inference, it is divided into two groups. The first group consists of 4 pharmacists who independently perform visual identification on a dataset containing 10 medications. The other group is AI recognition. The results are presented as true and false. We utilize a chi-square test to investigate whether there is a significant difference between AI and visual recognition.
Results
Total 150 images are comprised the database. 120 images are used as training set and 30 images are used as validation set. The training accuracy is 96.774%. In terms of inference, the success rate of AI recognition is 97.5% (39 true and 1 false), while human visual recognition is only 42.5% (17 true and 23 false), indicating a significant difference (p<0.001).
Conclusion
AI deep learning significantly increases the success rate of recognizing unmarked tablets. Future goals should focus on expanding the database by increasing the number of data entries and improving the speed of database establishment.
The issues of drug omissions and incorrect drug identification are great important in transition of care. Pharmacists face challenges in visually identifying many unmarked and unlabeled tablets.
Objectives
This study aims to establish an image database using artificial intelligence (AI) deep learning to replace manual recognition by the human eye, especially for medications that are difficult to distinguish.
Methods
(1) Database establishment: We choose 10 pills (in supporting documents) and each medication tablet is photographed 15 images, capturing the front view and the other sides. The angles are not restricted. We also employ a ruler-backed platform, measured in centimeters to enhance the accuracy of deep learning. (2) Training and validation: We utilize EfficientNet-b5 as the training model and the model type is multi-class classification. Random-resized-crop, random-horizontal-flip, color-jitter and random-rotation are applied for data augmentations. Normalization of images is applied as well. (3) Inference: We use the Inferencer object for inference, obtains the results, and prints them in JSON format by Python. The packages include simplejson, numpy, torch and so on. All processes are coded using Python. (4) Result: The primary result of training is accuracy. In terms of inference, it is divided into two groups. The first group consists of 4 pharmacists who independently perform visual identification on a dataset containing 10 medications. The other group is AI recognition. The results are presented as true and false. We utilize a chi-square test to investigate whether there is a significant difference between AI and visual recognition.
Results
Total 150 images are comprised the database. 120 images are used as training set and 30 images are used as validation set. The training accuracy is 96.774%. In terms of inference, the success rate of AI recognition is 97.5% (39 true and 1 false), while human visual recognition is only 42.5% (17 true and 23 false), indicating a significant difference (p<0.001).
Conclusion
AI deep learning significantly increases the success rate of recognizing unmarked tablets. Future goals should focus on expanding the database by increasing the number of data entries and improving the speed of database establishment.