Evaluation of breast tumor morphologies from African American and Caucasian patients.

The primary aim of this research was to investigate potential differences of breast tumor morphologies across African American and Caucasian racial groups by utilizing machine learning (ML) and artificial intelligence (AI) methods. While breast cancer disparities can partially be attributed to social determinants of health, tumor biology also contributes to survival outcomes. The rate of breast tumor growth is largely dependent on the extracellular matrix (ECM). Current research suggests the cellular components of the ECM may vary among racial and ethnic populations, and this may contribute to the incidence of cancer in African Americans. We utilized a supervised AI method to evaluate morphological differences between African American and Caucasian breast cancer tumors. Images used for analysis were downloaded from the Cancer Genome Atlas (TCGA) biorepository stored in the NIH Genomic Data Commons (GDC) data portal. We designed an ML classifier using the AlexNet model provided in PyTorch's torchvision package. The model was pre-trained and adapted via transfer learning resulting in a classification accuracy of 92.1% when using our breast cancer tumor image database split into 80% of training set and 20% of testing set. We interpreted the results of the AlexNet and ResNet50 models using LIME and saliency mapping as the explainers. Based on the images from our bi-racial testing set, this study confirmed significant variations of tumor and ECM regions in the different racial groups evaluated. Based on these findings, further analysis and characterization may provide new insight into disparities associated with the incidence of breast cancer.

CellTrackVis: analyzing the performance of cell tracking algorithms.

Live-cell imaging is a common data acquisition technique used by biologists to analyze cell behavior. Since manually tracking cells in a video sequence is extremely time-consuming, many automatic algorithms have been developed in the last twenty years to accomplish the task. However, none of these algorithms can yet claim robust tracking performance at the varying of acquisition conditions (e.g., cell type, acquisition device, cell treatments). While many visualization tools exist to help with cell behavior analysis, there are no tools to help with the algorithm's validation. This paper proposes CellTrackVis, a new visualization tool for evaluating cell tracking algorithms. CellTrackVis allows comparing automatically generated cell tracks with ground truth data to help biologists select the best-suited algorithm for their experimented pipeline. Moreover, CellTackVis can be used as a debugging tool while developing a new cell tracking algorithm to investigate where, when, and why each tracking error occurred.