Analysing hidden vector representation in a convolutional neural network


Twan Bongers
Presentation Student Presentation
Date 2019-01-15
Time 16:00
Location Carré 3446

The goal of the individual assignment is to use unsupervised deep learning, in particular a Convolutional Autoencoder (CAE), to make a tool for identifying pelvic floor states of pregnant women.

Unsupervised deep learning is used to structure unlabelled data by finding discriminating features hidden in the data. This is done by using an autoencoding neural network. This network will encode and decode the image and tries to make the output image similar to the input image. Between this encoding and decoding of the image, a vector representation can be obtained from the network.

This vector is analyzed to find any relation in deeper layers of the neural network and its dimensions can be reduced by using dimensionality reduction methods to visualize the data in a lower dimension. The available dataset consists of ultrasound images from pregnant women from 12, 36 weeks and 6 months after delivery. Transperineal ultrasound is not harmful to patients and thus it is possible to create a large dataset of echo images from 258 women. The pictures are made in three different states of contraction: contracted, rest and valsalva. From these pictures, a doctor can see if the pelvic floor has any damage. The goal is to use a computer to find any clinically relevant information in these images and data of the 258 women. This is done according to the following research question: How do we use unsupervised deep learning to obtain clinical relevant information from pelvic floor ultrasound images?

From this obtained data, the computer can create a prediction for possible changes in functionality of the pelvic floor during/after pregnancy. Classification between the contraction and valsalva is possible using a convolutional autoencoder neural network and applying the t-SNE dimension reduction method. This classification is graphically displayed in a 2D scatter plot which can be manually analyzed.

Posted on Wednesday, November 14, 2018