Magnetic resonance imaging (MRI) is the primary modality for the imaging of soft tissues (e.g. brain, muscle, liver). Therefore, it is an essential radiological tool for diagnosis and surgical planning. The contrast in MR images is due to tissues responding differently to the magnetic fields generated by the scan-ning system. This response can be de-scribed by the physical properties of the tissue (e.g. proton density, magnetic relaxation) and the magnetic fields. These physical properties are represent-ed by multiple parameters that can be estimated through quantitative MRI (qMRI) methods. The parameters are considered more reproducible than con-ventional MR images, which simplifies the comparison of MR data from differ-ent subjects or scanning systems. Esti-mating multiple parameters simultane-ously is needed to reduce error from system imperfections and deliver accu-rate estimates of the physical tissue pa-rameters...@en

This thesis explores a novel optical architecture for Whole Slide Imaging (WSI). This new architecture allows for multi-focal (3D) image acquisitions in a single scan pass. The multi-focal imaging capability is used to demonstrate 3D phase imaging and 3D imaging of thick tissue sections on a prototype scanner. Further, instrumentation for the extension of WSI to fluorescence imaging is developed: a technologically robust and cost effective method based on LEDs and a highly efficient method based on a multi-line laser illumination source. A WSI system is an optical instrument aimed at creating digital images of bio- logical samples mounted on a microscopy slide at a high throughput. WSI systems image tissues over large fields of view (∼ few cm), in 2D or in 3D (up to hundred layers of μm thickness), and at cellular resolution (∼1 μm). They are applied in high throughput screening in biology, and for novel computer aided medical diagnoses in the field of digital pathology. The optical architecture explored in this thesis is based on a tilted multi-line image sensor concept, originating from Philips. The goal of multi-line image ac- quisition is to enable closed-loop autofocus scanning, but it also allows for multi- focal image acquisitions. The core of this scanner concept lies in a novel design for a multi-line image sensor. The sensor is experimentally characterized for gain and noise, and a system model is developed to find the optimum signal-to-noise ratio (SNR) given the available photo-electron flux. This showed that images with a very high SNR of 292 can be acquired, provided that a sufficiently high photon flux can be realized. Two major issues with the sensor were found in our exper- imental characterization. At high line rates, the sensor showed missing symbols, leading to non-linearities in the read out. Further, the sensor showed high frequent fluctuations in the gain. 3D phase imaging and 3D imaging of thick specimens are two novel contrast modalities based on computational imaging techniques and are enabled by the availability of multi-focal images. For both techniques simplified algorithms were developed compatible with parallel processing at very high speeds. For 3D imaging of thick specimens a deconvolution technique is developed for improving the in- herently low axial contrast. 3D phase imaging is realized by a simplified algorithm for Quantitative Phase Tomography (QPT). QPT imaging is found to be able to im- age the sites labeled for Fluorescence in situ Hybridization (FISH) imaging, and provide additional structural information on unlabeled tissues or tissues stained for immunofluorescence. A system design study is presented showing that the in- plane transfer function has the character of a band-pass spatial frequency filter. The major opportunity for WSI systems to become compatible with fluores- cence imaging is addressed by the development of two imaging modalities. First, a widefield fluorescence WSI system with an LED illumination source is developed and built. A color sequential illumination strategy in combination with multi-band dichroics is used for multi-color imaging using a single monochromatic sensor. The main speed limitation is formed by the exposure time required to capture enough photo-electrons for a decent SNR. Based on the experimental results, a system with 96 Time Delayed Integration (TDI) lines is estimated to achieve a rea- sonable throughput of about 130 kPixel/s. This makes scanning possible of an area of 15 × 15 mm2 in three colors in about 23 min. Second, a novel optical architecture for multi-focal fluorescence image acqui- sitions based on a laser illumination source is proposed and realized in a proto- type. Illumination PSF engineering using diffractive optics is applied to generate a set of parallel scan lines in object space, that span a plane conjugate to a tilted im- age sensor. An important new element in the design is the use of higher order astig- matism to improve the uniformity of peak intensity and line width along the scan lines. Focusing the illumination on the sample provides a very high illumination efficiency and a confocal suppression of background. This optical architecture is projected to ultimately achieve a throughput of several hundreds MPixel/s, which would enable scanning an area of 15 × 15 mm2 in 8 layers in less than a minute. This thesis is concluded with an outlook to opportunities for future research in WSI systems. The challenges and some potential solutions for using a general purpose scientific CMOS (sCMOS) camera for multi-line scanning of a tilted ob- ject plane, and some opportunities for extension of WSI techniques to Light Sheet Microscopy (LSM) and Structured Illumination Microscopy (SIM) are discussed. In summary, this thesis investigates the imaging qualities and extension to computational imaging modalities of a brightfield WSI system and describes two approaches for fluorescence WSI. @en

The millions of filter weights in Convolutional Neural Networks (CNNs), all have a well-defined and analytical expression for the partial derivative to the loss function. Therefor these weights can be learned from data with a technique called gradient descent optimization. While the filter weights have a well-defined derivative, the filter size has not. There is currently no other way to optimize filter sizes, then with an exhaustive search over multiple CNNs trained with different filter sizes.
In this report, we propose a new filter called Structured Receptive Field of which the filter size can be optimized during the training stage. This new filter constitutes a parameterization of a normal filter as a linear combination of all 2D Gaussian derivatives up to a certain order. Instead of learning the weights of the resulting filter directly, we learn the weights of the linear combination and the sigma of the Gaussian derivatives that implicitly define the filter weights. The advantage of parameterizing normal filters in this way, is that gradient descent optimization of the continuous sigma parameter which has a well-defined derivative, can be used as a proxy for optimizing the discrete filter size which has no well-defined derivative.
The basic idea for this parameterization comes from scale-space theory in which the scale of image features is studied by parameterizing the image features with a continuous scale parameter. Thereby explicitly decoupling the spatial structure of image feature from the scale at which it occurs in the image. Structured receptive fields have several compelling advantages: they can learn bigger filters without increasing the number of learnable parameters and without increasing the complexity of the structure of the filter.
In this report, we both provide theoretical and empirical evidence that structured receptive fields can approximate any filter learned by normal CNNs. I.e. we show that structured receptive fields of order 4 can approximate the filters in all layers of a pre-trained AlexNet. Furthermore, we demonstrate empirically that structured receptive fields can indeed learn their own filter size during training, and that this extra ability over normal filters seems to give them an advantage over normal filters when used for classification tasks. I.e. by replacing normal filters in a DenseNet architecture and keeping every- thing else the same, a ~1% higher test accuracy was obtained on the highly competitive CIFAR-10 benchmark dataset.
It would not be wise to draw hard conclusions from a single training run on a single dataset, but this result definitely looks promising. In future research, we hope to demonstrate that, because of their extra ability to learn their filter size, replacing normal filters with structured receptive fields will always lead to strictly better or equal performance.