Bio

Alexandra Berg is an undergraduate student majoring in Computation and Cognition at Massachusetts Institute of Technology. Her early experience at the Nordic Institute of Theoretical Physics (NORDITA) and MIT’s Plasma Science and Fusion Center (PSFC) saw her researching and developing magnetohydrodynamic flow models. However, her interest in analytics and passion for healthcare has since prompted her to participate in several healthcare-oriented research and startup ventures. When not in front of a computer, Alex enjoys reading, painting, and backpacking.

Project: Image Segmentation of Histopathology Slides

Goal

Histopathological examination is a key means of diagnosing cancers. One particularly pertinent task within digital histopathology is image segmentation: the delineation of cell entities. Current segmentation models are largely based on handcrafted algorithms that require extensive feature engineering; in addition, they are typically only capable of segmenting a single type of cell entity. These factors currently prohibit their deployment in clinical settings. Thus, the aim of this project is to develop a model that is capable of performing segmentation across multiple types of cell entities without relying on handcrafted features.

Summary of Results

The project has utilized three copies of the Ademxapp A1 model pre-trained on the Cityscape Dataset and trained them to segment one type of cell entity each. The three networks were trained using publicly available datasets and, unlike earlier nets, did not rely on extensive sets of handcrafted features. The first model was trained to segment tubules and achieved a pixelwise accuracy of 89.3%; the second model was trained to segment epithelial cells and achieved an accuracy of 89.2%; finally, the third model was trained to segment nuclei and achieved an accuracy of 93.1%.

Future Work

Future work will first feature an attempt at increasing the accuracy of the three individual nets by implementing a mean-frequency balancing layer. Later on, an attempt will be made at creating a net that utilizes this mean-frequency balancing layer and a combined training scheme featuring all the available dataset to successfully segment multiple tissue-, cell- and organelle-level entities at once.