Skip to content
Back to Projects

2018 Data Science Bowl

Instance segmentation challenge to detect cell nuclei in microscopy images for medical research. Top 12% placement in Kaggle competition.

Winner Top 12% (IOU: 0.41553) Kaggle
View on GitHub

Skills

Computer Vision Instance Segmentation Deep Learning Data Augmentation

Tools

Python Keras TensorFlow OpenCV

Instance segmentation competition focused on improving nuclei detection from electron microscopy images for medical discovery.

Challenge

Detected cell nuclei across diverse image types with varying quality, coloration, and cellular conditions. Only 670 images were available with significant labeling issues (15-20% incorrect or missing labels), requiring manual corrections and augmentation techniques.

Approach

Images came from heterogeneous electron microscopy technologies showing cells in various states (normal, apoptotic, karyorrhexic, with blebbing) from multiple species. Developed custom performance metrics and extensive data augmentation pipelines.

Technologies

Built with Keras and TensorFlow for deep learning, OpenCV for image processing, and standard Python data science stack (Pandas, NumPy, Matplotlib, Seaborn).