Identifying Polyploid Cells in Tissue Images to Aid in Drug Development for Diffuse Large B-Cell Lymphoma

C. Rouse
Southwest Research Institute,
United States

Keywords: cancer, diagnostics, artificial intelligence, segmentation


Chromosome instability (CIN) and aneuploidy are classical hallmarks of cancer, caused by gross genomic rearrangements and cell cycle slippage. In a tumor, CIN is variable from cell to cell and can include changes in chromosome structure, chromosome number, and chromosomal rearrangements. Aneuploidy is a direct outcome of this instability. Polyploidy is a form of aneuploidy caused by whole genome doubling, resulting in several nuclei within the same cell. While rare in normal human tissues, polyploidy has been found in almost all cancers and is a known factor of therapeutic resistance. Currently, tumor biopsies must be sliced, stained, imaged, and assessed by eye. The high degree of similarity between polyploid cancer cells and surrounding cancer cells make it difficult to accurately identify polyploid cells both clinically and in research. Automating the identification of polyploid cells using Artificial Intelligence (AI) technology would aid drug development and research efforts targeting polyploidy. With further research, the use of AI polyploidy identification in patient samples could allow clinicians to use polyploid severity as a predictive marker of therapeutic resistance and modify therapeutic regimen as necessary. The hypothesis for this research is that the application of computer vision and neural networks for the task of instance-aware semantic segmentation, combined with domain expertise from medical professionals, can automate the process of identifying and classifying polyploid cells in tissue samples. Convolutional neural networks are known to be the gold standard for computer vision tasks in terms of accuracy and efficiency. Thus, they were fundamental to the polyploid quantification algorithm. More specifically, we used a custom convolutional neural network for instance-aware semantic segmentation, which has the capability to learn to identify both the location and area of different classes of objects. Applying instance-aware semantic segmentation is a popular approach for quantifying the contents of images because it accounts for spatial patterns among neighboring pixels. Our custom machine learning algorithm performed instance-aware semantic segmentation to identify polyploid cells in tissue images. The cells were labeled as one of the three classes. The 30 labeled images were used to train the instance-aware semantic segmentation algorithm. We applied a stochastic gradient descent optimizer to solve for the model parameters by iterating over the dataset for 300 epochs, selecting subregions from each image via an affine transform augmentation. An additional 25 images from the same set of mouse liver data comprised the test dataset. The well-trained instance-aware semantic segmentation algorithm was able to associate certain pixel information with each type of polyploid cell. In images of mouse liver tissue samples obtained from the University of Texas Health Science Center at San Antonio, the algorithm predicted polyploidy in cells with 94.4% accuracy and predicted sub-classes of polyploidy with 86.7% accuracy. The algorithm was trained and tested on healthy mouse liver cells because they are easily available and naturally contain unharmful polyploid cells. Current efforts are focusing on identifying polyploidy in lymphoma biopsies and assessing how well a drug in the development stage can prevent polyploidy when administered with chemotherapy.