HOME /

AI for Clinical Decision Support in Cancer Screening

Objective

The objective of this project is to train AI to check human interpretations of cancer screening images.

Motivation

The focus of this project is to improve patient outcomes by training AI to analyze diagnoses differences between clinicians and an AI method. The goal is to build an AI model in a more decision support role, rather than planning for complete algorithmic independence to analyze whether this can therefore maximize uptake by clinicians, payers, health systems, and patients.

Experiment

Our proposed test case is in breast cancer screening, where US radiologists have an average sensitivity of 84% and specificity of 91%. Breast cancer screening is widely endorsed by clinicians and patients, but the costs associated with radiologist labor and false positives reduce its value. For these reasons, breast cancer screening is a good fit for the application of AI. We have identified several candidate datasets:
 

Each of these datasets contains mammograms and clinical annotations.

Skills Required

This project primarily requires familiarity with machine learning, deep learning. You will take on the responsibility of carefully pre-processing data and linking different databases. Knowledge of SQL and other database management is a plus for this part of the project. Next, you will train deep learning models for classification using imaging datasets. A foundational understanding of Machine Learning, basic statistics, and probability is required for this project. This includes familiarity with state-of-the-art Deep Learning (specifically models used for prediction and diagnosis using imaging data) and working with Python (NumPy, scipy, pandas, sklearn). Knowledge of Deep learning packages like Pytorch, Tensorflow, JAX is required. Familiarity with working on Harvard FAS Clusters and with GPUs is a plus.

The desired end goal is a publication analyzing whether AI can reliably detect and therefore act as a double-check to human experts and analyze the variability in AI’s performance. When applying, please include a CV, any relevant projects you may have worked on, your availability and commitment over Fall 2021, as well as a brief paragraph (a couple of lines) on why you would be a good fit for pursuing this project.

Note: Compensation for the project is set for 15$ per hour for 20 hours per week. We anticipate the project to take about 3-4 months to the coding and results-complete stage of the project. Of course you will be an author on any ensuing publications and artifacts of the project.

To apply, please contact Shalmali Joshi at shalmali@seas.harvard.edu with the subject: Application for AI for clinical decision support in cancer screening. We will review on a rolling basis until a suitable collaborator is found.