136879 research outputs found
Sort by
Using Raman Spectroscopy to Differentiate Between Various Genospecies of Borrelia in Mouse Blood
Get PDFLyme disease (LD) is a tick-borne illness caused by a number of genospecies of the Borreliella burgdorferi sensu lato complex. LD is notoriously difficult to diagnose both by serological (the reference standard) and molecular (e.g., PCR) assays because the LD pathogen is only transiently present in the bloodstream with a follow-up dissemination to collagen-rich and often avascular tissues (e.g., joints). We hypothesized that Bb could cause substantial changes in blood biochemistry that can be detected and identified by Raman spectroscopy (RS). Our findings showed that RS, in combination with Partial Least-Squares Discriminant Analysis (PLS-DA), can differentiate blood samples taken at various stages of mouse infection. We also found that our innovative approach can differentiate between uninfected (control) and infected mice with 90% accuracy, as well as between the three main pathogenic genospecies, B. afzelli, B. burgdorferi sensu stricto, and B. garinii with 85% accuracy. These results indicate that RS in combination with PLS-DA may transform clinical approaches for pathogen diagnostics in the nearest future enabling inexpensive, non-invasive, non-destructive screening for LD and other bacterial infections
Training Deep Learning Models for Biological Image and Signal Analysis Using Limited Data Sets
Get PDFMedical decisions often require highly detailed expert analysis of patient biological signals and imaging studies. Data deluge can impede this decision making but can be mitigated by using computer aided diagnostic (CAD) tools. Neural networks (NNs) have been incorporated into CAD apps, since they have achieved state of the art performance on a number of tasks in non-biological spaces. However, generalizability is still a challenge in biological tasks. NNs bias towards their training data. Heterogeneity in biological targets can cause NN failure modes due to mismatch between training and real-world data. The trend in DL literature has been to focus on increasing model capacity and data size. Leveraging longer train times and more computational resources. Other literature focuses on searching for difficult samples, removing uninformative data from the training set, and on data relationships that lead to inference. This disconnect suggests that sources for DL model generalizability are not yet fully understood.
The primary goal of this thesis is to understand factors attributed to model generalizability, especially in biological datasets. The studied tasks include image segmentation, physical parameter estimation, signal denoising, and deconvolution. The variation in tasks is needed to minimize the chance that our observations are due to specific quirks of any one set.
We start by studying the use of color spaces, and wavelet like features in a segmentation task. Since model architecture was the most relevant driver of segmentation performance, we investigate whether this is due to specific network sub structures. But randomly generated models varied by task and did not always outperform known models, suggesting that the data itself is the primary driver of performance. The question becomes whether removing details from the training data forces models to learn root aspects of the data. Data augmentation worked better than training on abstract data though. So, there���s more of a case for data coverage, which we briefly cover in the last part of this thesis