Oscan-AI: Automated Diabetic Retinopathy Grading & Lesion Detection
Project Title: Oscan-AI: Automated Diabetic Retinopathy Grading & Lesion Detection 1. Project Overview Data Type: DICOM Medical Imaging & Image (High-resolution Fundus Photography) Labeling Types: Object Detection, NER Classification (Medical), and Data Collection. 2. Project Scope The scope involves the preparation of a high-quality dataset consisting of fundus images and DICOM metadata to train a deep-learning model. The goal is to detect early signs of diabetic retinopathy, specifically identifying microaneurysms, hemorrhages, and exudates, while also classifying the severity of the disease based on clinical guidelines. 3. Specific Data Labeling Tasks Performed Object Detection (Bounding Boxes & Polygons): Annotators identify and outline specific pathological features such as: Microaneurysms: Small red dots (earliest sign). Hard Exudates: Yellowish flecks (lipid deposits). Cotton Wool Spots: White, fluffy patches (nerve fiber damage). NER Classification (Metadata): Extracting and classifying structured data from DICOM headers and associated clinical notes, such as: Patient Demographics: Age, Gender. Technical Parameters: View position, pixel spacing, and device manufacturer. Image-Level Classification: Assigning a global severity grade (0–4) to each image based on the International Clinical Diabetic Retinopathy (ICDR) scale: 0: No DR 1: Mild Non-proliferative DR (NPDR) 2: Moderate NPDR 3: Severe NPDR 4: Proliferative DR (PDR) 4. Project Size Total Dataset:50,000 unique retinal images. Source: Multi-center clinical data (Europe and Southeast Asia) to ensure ethnic diversity in retinal pigmentation. Duration: 6 months. 5. Quality Measures To ensure the high precision required for medical diagnostics, the following Quality Assurance (QA) protocols are implemented: Gold Standard Consensus: A subset of 10% of the data is pre-annotated by Senior Ophthalmologists. Regular annotators must maintain a >95% Intersection over Union (IoU) match against this "Gold Standard" to remain on the project. Triple-Pass Verification: Each image is labeled by two independent annotators. If their labels disagree (e.g., one marks a lesion the other missed), a third expert "arbiter" (a certified retinal specialist) performs the final adjudication. Inter-Rater Reliability (IRR): We utilize Cohen’s Kappa coefficient to measure the level of agreement between annotators for the 0–4 severity grading, targeting a score of 0.85 or higher. DICOM De-identification Audit: A mandatory automated and manual check to ensure all Protected Health Information (PHI) is stripped from the metadata before the training phase begins.