Search

The application and comparison of U-Net convolutional neural network architectures is proposed in this work to guarantee a fast and accurate convergence of inverse scattering problems solved by Born iterative method, even in the presence of strong scatterers. Starting from a preliminary configuration proposed by the authors in some recent papers, two variants are introduced and discussed to significantly reduce the computational cost, while guaranteeing convergence with very high accuracy in the dielectric profiles reconstruction when considering strong scatterers, such as tumors, thus working as a regularization process to mitigate the induced non-linearity. As a further enhancement, a novel approach is introduced which integrates U-Net and Resnet models to realize a segmentation process, thus leading to the effective feature extraction and the accurate identification of anomalies within healthy tissue. Numerical assessments on a variety of breast models including abnormal lesions are discussed to successfully validate the proposed machine learning approach, through the adoption of properly defined evaluation metrics.

Wild-type transthyretin amyloidosis (ATTRwt) is an underdiagnosed and potentially fatal disease. Interestingly, ATTRwt deposits have been found to deposit in the ligamentum flavum (LF) of patients with lumbar spinal stenosis before the development of systemic and cardiac amyloidosis. In order to study this phenomenon and its possible relationship with LF thickening and systemic amyloidosis, a precise method of quantifying amyloid deposits in histological slides of LF is critical. However, such a method is currently unavailable. Here, we present a machine learning quantification method with Trainable Weka Segmentation (TWS) to assess amyloid deposition in histological slides of LF. Images of ligamentum flavum specimens stained with Congo red are obtained from spinal stenosis patients undergoing laminectomies and confirmed to be positive for ATTRwt. Amyloid deposits in these specimens are classified and quantified by TWS through training the algorithm via user-directed annotations on images of LF. TWS can also be automated through exposure to a set of training images with user-directed annotations, and then applied] to a set of new images without additional annotations. Additional methods of color thresholding and manual segmentation are also used on these images for comparison to TWS. We develop the use of TWS in images of LF and demonstrate its potential for automated quantification. TWS is strongly correlated with manual segmentation in the training set of images with user-directed annotations (R = 0.98; p = 0.0033) as well as in the application set of images where TWS was automated (R = 0.94; p = 0.016). Color thresholding was weakly correlated with manual segmentation in the training set of images (R = 0.78; p = 0.12) and in the application set of images (R = 0.65; p = 0.23). TWS machine learning closely correlates with the gold-standard comparator of manual segmentation and outperforms the color thresholding method. This novel machine learning method to quantify amyloid deposition in histological slides of ligamentum flavum is a precise, objective, accessible, high throughput, and powerful tool that will hopefully pave the way towards future research and clinical applications.