“Detection of Pneumonia in Medical Images”
Budget: £70 – £100 GBP
You are required to use Machine Learning techniques to tackle the problem of “Detection of
Pneumonia in Medical Images”. According to NHS records (https://aqua.nhs.uk/wpcontent/uploads/2020/11/Respiratory-thematic-report-revised-20201130.pdf), there were 272
thousand hospitalisations of Pneumonia in England in 2019. In the USA, it is one of the top 10
causes of death (https://www.medicalnewstoday.com/articles/282929#heart-disease).
Diagnosing Pneumonia requires careful analyses of chest radiographs by highly trained
specialists exposed to large amounts of images every day. Solutions to automate early
diagnoses would help in diagnosing such a disease. This task consists of creating image
classifiers to predict whether there is pneumonia (see image on the right) or not (see image on
the left) in an input image.
The dataset used in this task is from the following Kaggle competition:
https://www.kaggle.com/c/rsna-pneumonia-detection-challenge
You are expected to explore a range of machine learning classifiers, inspired by the various
models and categories explored within the module and beyond (i.e. from reading and literature).
At least two of the deep learning classifiers discussed in the lectures and/or workshops should
be included as baselines. In addition, at least one of your proposed classifiers should attempt to
go beyond the module in terms of architectural, approach, and/or algorithmic details.
You will then investigate their performance, compare and critique them to justify your
recommended classifier(s). This should include metrics such as TP/FP rates, Precision-Recall,
F-measure, and any other relevant metrics. In this assignment you are free to train any
classifier, to do any pre-processing of the data, and to implement your own algorithm(s) instead
of only using libraries. While you are encouraged to make your own implementations, you can
use libraries (such as Tensorflow or Pytorch) to train your your deep neural networks. But you
should clearly mention your resources, acknowledge appropriately, and compare between
classifiers and their results in your report.
TASK 2:
You are required to use Machine Learning to tackle the problem of “Game Learning”. Your goal
in this task is to train Deep Reinforcement Learning (DRL) agents that receive image-inputs
from a game simulator, and that output game actions to play the game autonomously. The
following simulator will be used to play the game of SuperMarioBros 1-1-v0:
https://github.com/Kautenja/gym-super-mario-bros
You are required to use your knowledge acquired in the module regarding DRL agents, and
knowledge acquired from additional recommended readings. This will be useful for investigating
the performance of those agents, and for comparing and criticising them so you can recommend
your best agent. You are expected to evaluate your agents using metrics sush as Avg. Reward,
Avg. Q-Value, Avg. Game Score, Avg. Steps Per Episode, and Training and Test Times.
You are expected to train at least three different agents (in addition to any baseline provided in
the module), which can differ in their state representation (CNN, CNN-RNN, CNN-Transformer)
and/or different learning algorithms. Once you have decided the agents that you want to report,
you should train them with three different seeds and average their results. If you report learning
curves, they should be based on those average results instead of using a single seed (run). You
are expected to justify your choices in terms of architectures, hyperparameters and algorithms.
In this assignment, you are free to train any DRL agent, in any programming language, to preprocess the data, and to implement your own solutions whenever possible. While you are free to
use libraries, you should not use fully available solutions. So please mention your resources
used, acknowledge appropriately, and compare between agents in your report.
Pneumonia in Medical Images”. According to NHS records (https://aqua.nhs.uk/wpcontent/uploads/2020/11/Respiratory-thematic-report-revised-20201130.pdf), there were 272
thousand hospitalisations of Pneumonia in England in 2019. In the USA, it is one of the top 10
causes of death (https://www.medicalnewstoday.com/articles/282929#heart-disease).
Diagnosing Pneumonia requires careful analyses of chest radiographs by highly trained
specialists exposed to large amounts of images every day. Solutions to automate early
diagnoses would help in diagnosing such a disease. This task consists of creating image
classifiers to predict whether there is pneumonia (see image on the right) or not (see image on
the left) in an input image.
The dataset used in this task is from the following Kaggle competition:
https://www.kaggle.com/c/rsna-pneumonia-detection-challenge
You are expected to explore a range of machine learning classifiers, inspired by the various
models and categories explored within the module and beyond (i.e. from reading and literature).
At least two of the deep learning classifiers discussed in the lectures and/or workshops should
be included as baselines. In addition, at least one of your proposed classifiers should attempt to
go beyond the module in terms of architectural, approach, and/or algorithmic details.
You will then investigate their performance, compare and critique them to justify your
recommended classifier(s). This should include metrics such as TP/FP rates, Precision-Recall,
F-measure, and any other relevant metrics. In this assignment you are free to train any
classifier, to do any pre-processing of the data, and to implement your own algorithm(s) instead
of only using libraries. While you are encouraged to make your own implementations, you can
use libraries (such as Tensorflow or Pytorch) to train your your deep neural networks. But you
should clearly mention your resources, acknowledge appropriately, and compare between
classifiers and their results in your report.
TASK 2:
You are required to use Machine Learning to tackle the problem of “Game Learning”. Your goal
in this task is to train Deep Reinforcement Learning (DRL) agents that receive image-inputs
from a game simulator, and that output game actions to play the game autonomously. The
following simulator will be used to play the game of SuperMarioBros 1-1-v0:
https://github.com/Kautenja/gym-super-mario-bros
You are required to use your knowledge acquired in the module regarding DRL agents, and
knowledge acquired from additional recommended readings. This will be useful for investigating
the performance of those agents, and for comparing and criticising them so you can recommend
your best agent. You are expected to evaluate your agents using metrics sush as Avg. Reward,
Avg. Q-Value, Avg. Game Score, Avg. Steps Per Episode, and Training and Test Times.
You are expected to train at least three different agents (in addition to any baseline provided in
the module), which can differ in their state representation (CNN, CNN-RNN, CNN-Transformer)
and/or different learning algorithms. Once you have decided the agents that you want to report,
you should train them with three different seeds and average their results. If you report learning
curves, they should be based on those average results instead of using a single seed (run). You
are expected to justify your choices in terms of architectures, hyperparameters and algorithms.
In this assignment, you are free to train any DRL agent, in any programming language, to preprocess the data, and to implement your own solutions whenever possible. While you are free to
use libraries, you should not use fully available solutions. So please mention your resources
used, acknowledge appropriately, and compare between agents in your report.