An end-to-end machine learning computer vision pipeline designed to digitize handwritten medical prescriptions into structured, machine-readable records.
Key Highlights
- Deep Learning Recognition: Integrated Microsoft TrOCR (Transformer-based Optical Character Recognition) fine-tuned for handwritten doctor prescriptions.
- Computer Vision Preprocessing: Implemented adaptive thresholding, deskewing, denoising, morphological operations, and line segmentation to maximize OCR accuracy.
- Entity Extraction & Matching: Automated extraction of drug dosage, frequency, and duration with fuzzy matching against standard pharmaceutical drug databases.
- Confidence Scoring & Flagging: Designed an uncertainty scoring algorithm to flag ambiguous handwriting for manual human pharmacist review.
- Service Architecture: Packaged as a high-performance RESTful API service using FastAPI, complete with interactive OpenAPI documentation and Docker containerization.