Image Processing For Healthcare Industry

About the Client

The client is a healthcare company specializing in human reproduction and embryology. They faced the challenge of determining the success of insemination in the early stages based on videos of embryos.

Problem

The specific challenge was to classify videos of embryos into three groups: Liveborn (embryo will result in a live birth), No implantation (embryo will not implant), and Miscarriage (embryo will result in a miscarriage). This classification would help doctors in deciding which embryo had the highest probability of resulting in a successful pregnancy.

Solution

The SmartCat team addressed the problem by implementing a machine learning model using image processing techniques and deep learning algorithms.

1. Preprocessing: The team converted the video files into a set of images (frames) using OpenCV. Various transformations were applied to extract embryos, and data augmentation techniques were used to increase the training set.
2. Dimensionality Reduction (Embedding): The team employed unsupervised learning techniques, specifically Variational Autoencoder and Stacked Denoising Autoencoder, to extract features from the image frames and create a lower-dimensional representation as vector embeddings.
3. Video Classification: The embeddings were used as sequences to make final predictions for the video. Support Vector Machine (SVM) was used for single image classification, and Long Short Term Memory (LSTM) was used for whole video classification.

Results

The approach achieved solid results in classifying images and videos on the test set. The next step for the team is to extend the model with a larger video dataset and combine results from images and metadata to improve accuracy further.

Smart Tip

In similar situations, it is important to consider using unsupervised learning techniques for dimensionality reduction and feature extraction, especially when dealing with video data. Data augmentation can also help expand the training set and improve model performance.

Technologies Used

• Python
• TensorFlow
• Deep Learning
• Neural Networks
• OpenCV