Stay Informed!

Sign up for one
of our newsletters.

Quin Xie – Research Studentship – 2018

Quin Xie is an Undergraduate Science Student at the University of Toronto.

Quin’s project has been generously supported by a gift from the Taite Boomer Foundation

About the research

Project title: “Automated histopathologic classification of brain tumours”

Brain tumours consist of numerous heterogeneous diseases with immensely variable therapies and outcomes. Analyzing pathological slides is essential for predicting prognosis of the tumour. Despite years of training and practice of expertise for pathological analyses, pathologists may still miss subtle microscopic features of diseases, or fail to reach agreement based on interpretation of slides. Recent advances in molecular analyses allow definitive differentiation between tumour types. However, testing results may take days to weeks, limiting their utility for guiding clinical decision making in acute and primary care settings. Clinical decisions thus still largely rely on microscopic findings.

This project exploits deep learning, a type of artificial intelligence specializing in pattern recognition tasks, to identify different tumour types based on microscopic features. Training of a convolutional neural network(CNN) model is accomplished by inputting large numbers of images with known clinical outcomes such as survival and therapeutic response to enable CNN to generalize microscopic patterns associated with corresponding clinical events. Since computers are better at incorporating more variables simultaneously when making decisions than human, it is anticipated that implementation of this approach will provide more accurate predictors of outcome and treatment response to patients.

About Quin, in her own words…

Quin Xie - Studentship - 2018Being awarded a Brain Tumour Research Studentship means that I can further develop my interests in brain tumour research in preparation for graduate studies. It was my interest in neuroscience that first drove me to study pathology.

Since my exposure to research as a volunteer in the Diamandis Lab, I have been developing a strong appreciation for the diverse research areas in brain tumours. Over the past year, I have been involved in the application of artificial intelligence to advance pathological analysis of brain tumours. Not only has the experience familiarized me with the research and clinical duties of a neuropathologist, but it has also deepened my understanding towards the potential of the brain tumour research for well-being of the patients.

This generous award will allow me to longitudinally continue my research in the Diamandis Lab and pursue my career in the field of brain tumour research. I anticipate that my contribution will facilitate the application of the research outcome and reduce cost and time for brain tumour treatment in the future. It is a great honour to receive a studentship and represent Brain Tumour Foundation of Canada during through my ongoing research in the Diamandis Lab.

Progress Update – September 2018

In this summer, we aimed to optimize the CNN model for brain tumour classification. However, we encountered two ineligible difficulties: 1) Despite the advantage of CNN in pattern recognition and feature extraction, 5 percent of total cases are wrongly classified; 2) Unlike pathologists who can give differential diagnoses or seek second opinions, the model is ignorant about the tumour subtypes not shown up in the training sets, and try to fit them into existent categories.

To address those problems, we employed a non-linear dimension reduction technique called t-distributed Stochastic Neighbor Embedding (t-SNE) It is able to show how CNNs organize histomorphologic information before the output layer by reducing all variables into 2 dimensions. To elaborate, we developed a quantitative and transparent approach to visualizing classification decisions prior to softmax compression. By discretizing the relationships between classes on the t-SNE plot, we superimposed randomly sampled regions of test images and use their distribution to render statistically-driven classifications. In addition to providing intuitive outputs for human review, this visual approach can carry out automated and objective multi-class classifications similar to more traditional and less-transparent categorical probability distribution scores. Importantly, this novel classification approach is driven by a priori statistically defined cutoffs. It therefore serves as a generalizable classification and anomaly detection tool less reliant on post-hoc tuning.

In the next summer, we hope to integrate our CNN model with other machine learning methods to incorporate other clinical variables, such as age, gender, molecular alterations, and survival data to correlate histopathologic information with patients’ treatment outcome. The report generated by our model will make the diagnoses more accessible to clinicians and patients, thus accelerate the process of personalization of medicine at the point of care.