Nan Wu: Deep Neural Networks Improve Radiologists’ Performance in Breast Cancer Screening
October is Breast Cancer Awareness Month, but CDS PhD student Nan Wu takes awareness into action by applying deep neural networks to breast cancer screening images.
Nan Wu did her undergraduate study at School for Gifted Young, University of Science and Technology of China where she graduated with a B.S. in Statistics and a B.A in Business Administration. The School for Gifted Young is designed for students two or three years younger than the average age for college students. “About half of the students were majoring in physics, and eighty percent are doing their PhD now,” Nan told us. “At the time, I wanted to be different, I didn’t want to be a PhD student or stay in academia so I chose to do a Masters. Along the way, I realized that research in data science actually is a field I would like to stay in.”
During Nan’s final year in the Master’s program at CDS, her attention was captivated by a project building models to make predictions based on breast cancer screening images: “At the time I met Professor Geras and Professor Cho they were recruiting students to join them in doing research for the breast cancer project they were working on. At the beginning I was working on the breast density classification task. I also focused on this task in my capstone project, collaborating with two other students. After that semester everything went well, and my first publication was accepted to be presented at a conference. Fortunately, I got the chance to start my PhD study and to continue my research project on the cancer classification task with our team. ”
The models created in this project delivered results with accuracy comparable to human radiologists, and the work has received great recognition: The paper was published at IEEE Transactions on Medical Imaging (TMI), and a photo from Nan’s paper is featured on an upcoming issue of Radiology, the most prestigious clinical journal on radiology in general.
But Nan is not stopping there. Part of the reason her passion lies in applying data science to health care is because there is an added level of challenge, as tasks related to health care are often more difficult than other deep learning problems. She is now taking her research to the next level by working on developing specific deep learning methods for mult-modal tasks:
“As humans, we have visual information and audio information and so on. Previously people were building models using a single source of input, but fusing multiple kinds of input together to get better performance is kind of a new problem. With breast cancer, we used a screening mammogram, but even in a screening mammogram there are multiple images from different views . Currently most of the deep learning models perform well when it is presented with one modality. However, it’s difficult to tell the model how to learn and combine information to make a decision, so that’s the goal of my current project.”
Another reason Nan enjoys working with data in healthcare is because of the way she can feel the real-world effects of her work. “It’s closer to our daily life,” she says. “You can see how your work is improving lives. I considered the detection of breast cancer for my research in applying deep learning models to medical imaging because breast cancer is the leading cause for women’s death in the United States. Since women over the age of 40 are encouraged to get annual screenings, we have collected a huge amount of medical imaging in an organized format. Further, many of these women return annually for their screening mammogram, allowing us to amass data in an organized format, to corroborate the reference standards with the electronic medical records, and to detect subtle imaging changes over time.”
Nan’s passion for data science and healthcare shine through whenever she speaks about either: “My research encourages me to learn more about cancer. While at the same time, it inspires me to diagnose problems in deep learning such as how doctors diagnose cancer,” she said. “We can develop metrics to observe learning dynamic as imaging techniques are used to observe cancer. Based on our observation, we can further propose new algorithms similar to the way you can develop new medicines to cure the kind of cancer.”
When people outside the data science community learn of the success projects such as Nan’s work with medical imaging processing have experienced, eyebrows and questions alike may rise about future purposes of such powerful deep learning tools. According to Nan, nobody is hoping to replace radiologists with models, but rather to augment their expertise with AI readouts. “With radiologists as domain experts, we can keep improving the model, and the model’s readout could be used as a second reader for them. By combining the AI and human power, we can only lead to a more reliable and convincing result for the patient.”
NYU Langone recently published an article on the work of Nan and her team. As data science continues to grow as a field, the power of AI is increasingly exchanging its sense of novelty in for a more widespread recognition of current and potential real-world applications. “I can see for example the NYU School of Medicine trying to embrace this change in AI related to the medical field, they are hiring more and more researchers. The requirements from medical field is enriching dimensions in data science research as well,” says Nan. “For example, when you are building a model, you should care not only about the performance but also interpretability since it has to be expandable to the radiologist and patient.”
By Mary Oliver
