Ensemble Learning for Medical Image Character Recognition based on Enhanced Lenet-5

Generally, Medical Imaging Modalities (MIM) have a distinctive nature of low contrast, complex background, and low resolution, containing burned-in textual data of patients. The conventional OCRs hardly recognise these burned-in textual data under these conditions, as they are designed for mainly bi-level text with a minimum resolution of 300 dpi. With a focus on solving these challenges, an enhanced CNN model for medical image character recognition (MICR) is proposed in this paper. The Lenet-5 architecture inspires this proposed Model. To further enhance this new technique to recognise visually similar characters, this paper proposes an ensemble classifier of CNN base learners. Intensive experiments are done using an open-source medical imaging dataset. The problem of low resolution at 96dpi and background interference is targeted by using small 3 X 3 CNN filters to extract local features and changing the pooling layer to a learning layer by replacing it with 5 X 5 filters with a stride of 2 and training on a low-resolution character dataset. The final prediction is based on a majority voting algorithm. The consensus of the base learners improves the model’s stability in recognising visually similar characters. Finally, our proposed models and the Lenet-5 are compared using the Medpix medical image collection. Further investigation shows that our proposed model shows a 10% increase in accuracy compared with the base model and other past algorithms in recognising burned-in textual data on medical imaging modalities.