X-Ray Classification Convolutional Neural Network

Our CNN model classifies a given chest x-ray as normal lungs, or lungs with pneumonia. The difference between a normal scan and a pneumonia scan is extremely subtle.

Comment 10

A link to our Google Doc writeup is included below, and here: https://docs.google.com/document/d/1IaX5Hk7FmZxB5qohdWGZ6urAFRdATEqDZXGnj3KVCKQ/edit?usp=sharing

Github link (also linked below): https://github.com/tucker-weed/cs1470_final-X-ray-CNN

Dataset link: https://www.kaggle.com/paultimothymooney/chest-xray-pneumonia

For a reference of how much our project has changed, here is our original 'About' section: Music interests are inherently sequential. You listen to one song, then you listen to another. We will implement a Transformer that, given a user's listening history, predicts the next song they will listen to. We will train our model using the LastFM-1K dataset, which contains the entire listening history of 1,000 users. This dataset, however, contains over 180,000 artists and 20 million individual events. Before feeding sequential data into our Transformer, we will first need to reduce the number of songs from over 1/2 million to a set of a few hundred from which we will recommend. We will do this by feeding concatenated listening history embeddings (not considering it as a sequence) to a simple set of linear layers. We will then feed this reduced number of songs into our Transformer model, and consider the level of success based on the distance between the embedding of the recommended song at time t and the embedding of the label.

We have two main sources on which we are basing our architecture. The first is the paper Deep Neural Networks for Youtube Recommendations published in 2016 (before Transformers, which were developed in 2017). The second is Behavior Sequence Transformer for E-commerce Recommendation in Alibaba published in 2019.

Ethics: Our dataset was collected from LastFM in 2010 and contains user history. Although this history is anonymized, this still may be a violation of privacy. Likely, upon signing up for LastFM, these people signed a privacy policy without reading it and unknowingly consented to having their listening history scraped. It contains fields such as the home country, date joined, and binary gender of the listener. If we decide to use these user features as inputs to our model, we may be subjecting ourselves to potential biases and preventing ourselves from valid applications to a broader user base that doesn't have these narrow characteristics.

The actual recommendation of music, in our view, poses less of a threat of societal harm than other forms of recommendations. For example, what a search engine chooses to put at the top of its results, or what appears at the top of your feed in social media, may be biased or divisive, and be broadly harmful to society. Music recommendations seem not to have as much leeway for damage. That said, we will be careful our model does not overlearn the preferences of a certain subset of listeners. However, given the restrictions of data in our dataset, this may be difficult.

Built With

Share this project:

Updates