A deep learning approach for multi-class classification of handwritten prescription images using CNN and Grad-CAM visualization
Author(s): Mohammed Nasih Ismael
Abstract: The identification of handwritten prescriptions is increasingly regarded as an important task during the digitization process occurring within healthcare. Written prescriptions are estimated to account for more than 13% of prescription errors. This research proposes a convolutional neural network (CNN)-based classification approach for the automatic classification of handwritten prescription drug images to a total of 78 classes. The dataset contains 60 images per class, 20 of which were used to train the network, 10 of which were used for validation, and the remaining 10 were used for testing. All images were resized to 64 × 64 pixels and converted to grayscale to provide input into the model. The CNN configuration achieved 80.38% validation accuracy and 69.74% testing accuracy. In order to interpret model outputs, Gradient-weighted Class Activation Mapping (Grad-CAM) was employed to highlight class-relevant areas of the images. The last component of this work evaluated and analyzed the model using confusion matrix analysis, accuracy per class, and misclassification patterns. Ultimately, this work indicates that the proposed method is feasible for the recognition of handwritten prescription and provides new information for the future practical application of this task in medical systems. This initial work is an important contribution to the growing area of AI-assisted medicine, providing an interpretable and scalable solution to drug identification.
Mohammed Nasih Ismael. A deep learning approach for multi-class classification of handwritten prescription images using CNN and Grad-CAM visualization. Int J Eng Comput Sci 2025;7(2):215-225. DOI: 10.33545/26633582.2025.v7.i2c.217
Research Journals