Team: 8
School: Clovis High Early College
Area of Science: Medicine
Interim: Definition of Problem:
Cancer treatment can often be a tedious process. Many times a patient must consult numerous health professionals to receive the best care and spend a prodigal amount of resources for human assistance. The use of neural networks to diagnose Cancer has taken a critical role in presenting the usefulness of the technology. We aim to utilize neural networks to compile the most desirable Cancer treatment plan; this includes treatment factors such as surgery, chemotherapy, and radiotherapy and is limited to the data available.
Computational Plan:
Python will be used to complete the coding. Neural network(s) will be created with the Keras API. Sequential and Functional models will be used. The Tkinter module will be utilized for GUI elements. Since our project Proposal, we believe it is best to not complete the Quantum Machine Learning (QML) version of our model. Reasons for this include a general unlikelihood of presenting supremacy with classical data, a technology that remains in its infancy, and a lack of resources (such as APIs) to effectively build this form of a model. We now intend to implement more sophisticated instruments into our classical model to improve the performance and advance our main objective more than would be possible with a QML version. At present, we plan to implement a multi-target regression with the Keras Functional API, improve our data preparation algorithm to combine datasets that contain the same variables, and do more to refine our model. We are also considering utilizing the weights and biases from our network to derive equations for the dependent variables.
Progress to Date:
So far, we have possessed four datasets that will be primarily used for training and testing the model. One of them contains imaging, radiation therapy, and clinical data from 627 head and neck squamous cell carcinoma (HNSCC) patients from MD Anderson Cancer Center. Another contains clinical data and CT scans from 137 HNSCC patients all treated by radiotherapy. More information about these and other datasets can be found in the README of the project GitHub. The early results of the model have been promising. In the “chemotherapy_given†variable in the HNSCC-HN1 dataset, a 92% training accuracy and an 80% testing accuracy have been achieved. The “cancer_surgery_performed†variable of the same dataset developed a 96% training accuracy and a virtually 100% testing accuracy. However, we have faced some uncertainty when including images. The model becomes much less accurate when images are inputted. Currently, the model converts the Dicom files (medical format) into jpeg images. Then, the jpeg images are digitized and put into a NumPy array. Some validation that this is an acceptable way of preprocessing images would be greatly appreciated.
Two members of our team are beginning to learn the functions of Python and how to operate the language. We hope to grow our understanding of Python. We plan on using Tkinter in Python to create our model. Our model is in the developing stages. The plan for the model is to learn the basics and have an understanding of how to create a GUI. After we have learned the basics, we will start building the structure of the model.
Expected Results:
After the development of our model, it could present the usefulness of machine learning in the medical field even further, beyond merely diagnosing a patient. More advanced versions of the concept could play a key role in the age of automation in which a machine is more suited for the role of a doctor than a human. Cancer treatment could be completed at a much more affordable cost and the result could be tuned precisely to the patient’s attributes and have a higher rate of success.
Citations:
1. How Treatment Is Planned and Scheduled: Cancer Treatment Scheduling. (n.d.). Retrieved November 23, 2020, from https://www.cancer.org/treatment/treatments-and-side-effects/planning-managing/planning-scheduling-treatment.html
2. Fchollet. (2020, April 04). The Functional API. Retrieved from https://keras.io/guides/functional_api/
3. Deriving Equation by Weights and Biases from a Neural Network. (n.d.). Retrieved from https://stackoverflow.com/questions/17510098/deriving-equation-by-weights-and-biases-from-a-neural-network/17523414
4. Brownlee, J. (2020, August 27). Deep Learning Models for Multi-Output Regression. Retrieved December 04, 2020, from https://machinelearningmastery.com/deep-learning-models-for-multi-output-regression/
5. The Keras functional API: Five simple examples. (2017, November 06). Retrieved December 04, 2020, from https://tomroth.com.au/keras/
Link to project GitHub: https://github.com/NM-Supercomputing-Challenge/Team8_Cancer_ML
Team Members:
Erynn Vetterly
Alexis Brandsma
Tristen Pool
Sponsoring Teacher: NA