Project Description

Kaggle's annual Data Science Bowl for 2018 was an instance segmentation competition aimed at improving the state-of-the-art in nuclei detection from electron microscopy images. My presentations on this competition are available in my repo.

Skills

Data Augmentation, Computer Vision, Image Processing, Instance Segmentation, Convolutional Neural Networks

Tools

Python, Pandas, NumPy, Keras, TensorFlow, Matplotlib, Seaborn, CV2, JSON

Motivation

Kaggle's annual Data Science Bowl for 2018 was a computer vision (instance segmentation) challenge; the mission was to write a better nuclei detector [for a variety of images generated by heterogeneous electron microscopy technologies] than the current state-of-the-art. In addition to evaluating models appropriate for this challenge, custom performance metrics were developed as Mean IOU is not available out of the box.

Images came in a variety of shapes and varying degrees of quality and coloration. Some were fluouresced, some grayscale, others were taken with dyes, some had artifacts on the slide that obscured portions of the cells, and so on. Additionally, the cells under consideration came from a wide variety of species and organisms. Cells were presented in various states: normal, apoptotic, karyorrhexic, no-to-severe blebbing, and so on.

To further compound the challenge, only 670 images were provided and labels on at least 15%-20% of the data was factually incorrect or missing entirely. Manual modification of label masks and data augmentation were required to generate accurate and sufficent data for final prediction.

See the complete project in my GitHub repository.