A Transfer Learning approach for handwritten drug names recognition

• Main Authors: Sehimi, Karim, Bendaoud, Fayssal, Djenandar, Mohammed Yacine, Benderbal, Hichem Haddou
• Format: Conference Proceeding
• Language: English
• Subjects: Adaptation models; Computational modeling; Convolutional Recurrent Neural Network; Handwriting recognition; Healthcare automation; Medical handwritting recognition; Process control; Recurrent neural networks; Training; Transfer learning
• ISSN: 2766-8665
• DOI: 10.1109/ICNSC58704.2023.10318974

In many developing countries, handwritten medical prescriptions remain a prevalent mode of documentation within healthcare systems. However, the legibility of these prescriptions often poses a significant challenge, leading to potential errors in patient care. To address this critical issue, our work focuses on developing a model for the recognition of medical handwritten prescriptions. The model architecture is composed of Convolutional Neural Networks (CNNs) and Bi-directional Long-short term memory (Bi-LSTMs) and a Connectionist temporal classification (CTC) Layer to compute loss between predicted sequences and ground truth sequences. However, the availability of annotated medical handwriting data is constrained, particularly due to privacy concerns. In response, we delve into transfer learning techniques to harness existing knowledge and enhance model performance. By leveraging transfer learning, we aim to optimize our recognition model using fine-tuning strategies. This study contributes to the field of medical handwriting recognition by exploring the potential of transfer learning in enhancing model accuracy and applicability within developing healthcare contexts.