Open Access

Srividya Tirunellai Rajamani

• This thesis explores deep learning based methods for health data, specifically on novel enhancements to attention mechanisms in diverse tasks of image and signal analysis. We demonstrate the effectiveness of our proposed attention mechanism enhancements in performance improvement, model complexity reduction, outlier detection as well as dealing with sparse and irregularly sampled time series data. In the context of medical image segmentation, effective handling of outliers is vital to ensure translation of research into clinical practise. Standard metrics used for reporting the performance of medical image segmentation algorithms report aggregate metrics across all patients. Due to this reporting, models that report superior performance could end up producing completely erroneous results, or even anatomically impossible results in a few challenging cases (corner-cases), albeit without being noticed. To counter this drawback, we propose a framework that helps to identify and reportThis thesis explores deep learning based methods for health data, specifically on novel enhancements to attention mechanisms in diverse tasks of image and signal analysis. We demonstrate the effectiveness of our proposed attention mechanism enhancements in performance improvement, model complexity reduction, outlier detection as well as dealing with sparse and irregularly sampled time series data. In the context of medical image segmentation, effective handling of outliers is vital to ensure translation of research into clinical practise. Standard metrics used for reporting the performance of medical image segmentation algorithms report aggregate metrics across all patients. Due to this reporting, models that report superior performance could end up producing completely erroneous results, or even anatomically impossible results in a few challenging cases (corner-cases), albeit without being noticed. To counter this drawback, we propose a framework that helps to identify and report corner cases. Further, we propose a novel balanced checkpointing scheme capable of finding a solution that has superior performance even on these corner cases. Deep neural networks with attention mechanism have shown promising results in many computer vision and medical image processing applications. One way to enhance attention is to build on the concept of deformability which was introduced in the context of convolutions. We propose a new Deformable Attention Network (DANet) that enables a more accurate contextual information computation in a similarly efficient way. Our novel technique is based on learning the deformation of the query, key and value attention feature maps in a continuous way. A deep segmentation network with this attention mechanism is able to capture attention from only the pertinent non-local locations. Deformability indicates that the attention mechanism could be further regularised. Hence we explore ways to regularise attention. We introduce a simple and low-overhead approach of adding noise to the attention block which we discover to be very effective when using an attention mechanism. Our proposed methodology of introducing regularisation in the attention block by adding noise makes the network more robust and resilient, especially in scenarios where there is limited training data. We incorporate this regularisation mechanism in the criss-cross attention block. This criss-cross attention block enhanced with regularisation is integrated in the bottleneck layer of a U-Net for the task of medical image segmentation. In the context of attention mechanism utilization in time-series data, we demonstrate the efficacy of using sparsely and irregularly sampled data when used in tandem with state-of-the-art existing attention based networks that are capable of handling sparse data. With our proposed sub-sampling approach, we demonstrate that time-series data could be further coarsely acquired. This could be of immense help for various applications where data acquisition and labeling is a significant challenge. By utilizing attention mechanisms in non-linear blocks in the context of GRU, we propose a novel Attention based GRU module. We demonstrate the effectiveness of this module to improve performance in the context of speech emotion recognition. Additionally, we also propose a novel metric for image quality assessment to compute the quality of a given image without a reference pristine quality image. Many of the image acquisition processes, especially in medical imaging, would immensely benefit from such a metric which can indicate if the quality of an image is improving or worsening based on adaptation of the acquisition parameters.…