MEDIREAD – AI System to Decode Doctors' Handwriting

MEDIREAD is an AI-powered system designed to recognize and extract text from doctors' handwritten prescriptions. The system leverages a Convolutional Neural Network (CNN) for handwriting recognition and integrates both local Tesseract OCR and the OCR.Space API for enhanced accuracy.

---

Features

• CNN-based handwriting recognition for prescriptions.
• Dual OCR processing system:
  • Local Tesseract OCR (offline, customizable)
  • OCR.Space API (high accuracy, cloud-based)
• Automatic preprocessing pipeline:
  • Grayscale conversion
  • Resizing to 64×64 pixels
  • Pixel normalization
  • Noise removal and thresholding
• Decision system to compare OCR results and select the most reliable output.
• JSON output with extracted prescription text and metadata.
• Visualization of uploaded prescriptions.

---

Dataset

• Source: Kaggle – Doctors Handwritten Prescription (BD Dataset)
• Format: Scanned prescription images (PNG, JPG, JPEG)
• Preprocessing:
  • Grayscale conversion
  • Resize to 64×64 pixels
  • Pixel normalization
  • Validation checks for corrupted/empty images
• Structure: Organized in folders by prescription types/handwriting styles

---

Model Architecture

• Framework: TensorFlow 2.x / Keras
• Type: Convolutional Neural Network (CNN)
• Input: 64×64×1 grayscale images
• Layers:
  • Conv2D → BatchNorm → MaxPooling2D → Dropout (3 blocks)
  • Flatten → Dense(256, ReLU) → BatchNorm → Dropout(0.5)
  • Output Layer: Softmax (multi-class) or Sigmoid (binary)
• Optimizer: Adam
• Loss Function: Categorical Crossentropy
• Training Accuracy: ~92% (sample dataset)
• Trained Model: prescription_model.h5