-
A study on the adequacy of common IQA measures for medical images
Authors:
Anna Breger,
Clemens Karner,
Ian Selby,
Janek Gröhl,
Sören Dittmer,
Edward Lilley,
Judith Babar,
Jake Beckford,
Timothy J Sadler,
Shahab Shahipasand,
Arthikkaa Thavakumar,
Michael Roberts,
Carola-Bibiane Schönlieb
Abstract:
Image quality assessment (IQA) is standard practice in the development stage of novel machine learning algorithms that operate on images. The most commonly used IQA measures have been developed and tested for natural images, but not in the medical setting. Reported inconsistencies arising in medical images are not surprising, as they have different properties than natural images. In this study, we…
▽ More
Image quality assessment (IQA) is standard practice in the development stage of novel machine learning algorithms that operate on images. The most commonly used IQA measures have been developed and tested for natural images, but not in the medical setting. Reported inconsistencies arising in medical images are not surprising, as they have different properties than natural images. In this study, we test the applicability of common IQA measures for medical image data by comparing their assessment to manually rated chest X-ray (5 experts) and photoacoustic image data (1 expert). Moreover, we include supplementary studies on grayscale natural images and accelerated brain MRI data. The results of all experiments show a similar outcome in line with previous findings for medical imaging: PSNR and SSIM in the default setting are in the lower range of the result list and HaarPSI outperforms the other tested measures in the overall performance. Also among the top performers in our medical experiments are the full reference measures DISTS, FSIM, LPIPS and MS-SSIM. Generally, the results on natural images yield considerably higher correlations, suggesting that the additional employment of tailored IQA measures for medical imaging algorithms is needed.
△ Less
Submitted 29 May, 2024;
originally announced May 2024.
-
A study of why we need to reassess full reference image quality assessment with medical images
Authors:
Anna Breger,
Ander Biguri,
Malena Sabaté Landman,
Ian Selby,
Nicole Amberg,
Elisabeth Brunner,
Janek Gröhl,
Sepideh Hatamikia,
Clemens Karner,
Lipeng Ning,
Sören Dittmer,
Michael Roberts,
AIX-COVNET Collaboration,
Carola-Bibiane Schönlieb
Abstract:
Image quality assessment (IQA) is not just indispensable in clinical practice to ensure high standards, but also in the development stage of novel algorithms that operate on medical images with reference data. This paper provides a structured and comprehensive collection of examples where the two most common full reference (FR) image quality measures prove to be unsuitable for the assessment of no…
▽ More
Image quality assessment (IQA) is not just indispensable in clinical practice to ensure high standards, but also in the development stage of novel algorithms that operate on medical images with reference data. This paper provides a structured and comprehensive collection of examples where the two most common full reference (FR) image quality measures prove to be unsuitable for the assessment of novel algorithms using different kinds of medical images, including real-world MRI, CT, OCT, X-Ray, digital pathology and photoacoustic imaging data. In particular, the FR-IQA measures PSNR and SSIM are known and tested for working successfully in many natural imaging tasks, but discrepancies in medical scenarios have been noted in the literature. Inconsistencies arising in medical images are not surprising, as they have very different properties than natural images which have not been targeted nor tested in the development of the mentioned measures, and therefore might imply wrong judgement of novel methods for medical images. Therefore, improvement is urgently needed in particular in this era of AI to increase explainability, reproducibility and generalizability in machine learning for medical imaging and beyond. On top of the pitfalls we will provide ideas for future research as well as suggesting guidelines for the usage of FR-IQA measures applied to medical images.
△ Less
Submitted 29 May, 2024;
originally announced May 2024.
-
Partially bonded crystals: a pathway to porosity and polymorphism
Authors:
Carina Karner,
Emanuela Bianchi
Abstract:
In recent years, experimental and theoretical investigations have shown that anisotropic colloids can self-organise into ordered porous monolayers, where the interplay of localised bonding sites, so called patches, with the particle's shape is responsible for driving the systems away from close-packing and towards porosity. Until now it has been assumed that patchy particles have to be fully bonde…
▽ More
In recent years, experimental and theoretical investigations have shown that anisotropic colloids can self-organise into ordered porous monolayers, where the interplay of localised bonding sites, so called patches, with the particle's shape is responsible for driving the systems away from close-packing and towards porosity. Until now it has been assumed that patchy particles have to be fully bonded with their neighbouring particles for crystals to form, and that, if full bonding cannot be achieved due to the choice of patch placement, disordered assemblies will form instead. In contrast, we show that by deliberately displacing the patches such that full bonding is disfavored, a different route to porous crystalline monolayers emerges, where geometric frustration and partial bonding are pivotal in the structure formation process. The resulting dangling bonds lead to the emergence of effectively chiral units which then act as building blocks for energetically equivalent crystal polymorphs.
△ Less
Submitted 2 May, 2024;
originally announced May 2024.
-
visClust: A visual clustering algorithm based on orthogonal projections
Authors:
Anna Breger,
Clemens Karner,
Martin Ehler
Abstract:
We present a novel clustering algorithm, visClust, that is based on lower dimensional data representations and visual interpretation. Thereto, we design a transformation that allows the data to be represented by a binary integer array enabling the use of image processing methods to select a partition. Qualitative and quantitative analyses measured in accuracy and an adjusted Rand-Index show that t…
▽ More
We present a novel clustering algorithm, visClust, that is based on lower dimensional data representations and visual interpretation. Thereto, we design a transformation that allows the data to be represented by a binary integer array enabling the use of image processing methods to select a partition. Qualitative and quantitative analyses measured in accuracy and an adjusted Rand-Index show that the algorithm performs well while requiring low runtime and RAM. We compare the results to 6 state-of-the-art algorithms with available code, confirming the quality of visClust by superior performance in most experiments. Moreover, the algorithm asks for just one obligatory input parameter while allowing optimization via optional parameters. The code is made available on GitHub and straightforward to use.
△ Less
Submitted 7 December, 2023; v1 submitted 7 November, 2022;
originally announced November 2022.
-
Limitations of neural network training due to numerical instability of backpropagation
Authors:
Clemens Karner,
Vladimir Kazeev,
Philipp Christian Petersen
Abstract:
We study the training of deep neural networks by gradient descent where floating-point arithmetic is used to compute the gradients. In this framework and under realistic assumptions, we demonstrate that it is highly unlikely to find ReLU neural networks that maintain, in the course of training with gradient descent, superlinearly many affine pieces with respect to their number of layers. In virtua…
▽ More
We study the training of deep neural networks by gradient descent where floating-point arithmetic is used to compute the gradients. In this framework and under realistic assumptions, we demonstrate that it is highly unlikely to find ReLU neural networks that maintain, in the course of training with gradient descent, superlinearly many affine pieces with respect to their number of layers. In virtually all approximation theoretical arguments that yield high-order polynomial rates of approximation, sequences of ReLU neural networks with exponentially many affine pieces compared to their numbers of layers are used. As a consequence, we conclude that approximating sequences of ReLU neural networks resulting from gradient descent in practice differ substantially from theoretically constructed sequences. The assumptions and the theoretical results are compared to a numerical study, which yields concurring results.
△ Less
Submitted 15 November, 2023; v1 submitted 3 October, 2022;
originally announced October 2022.
-
Hierarchical self-assembly of anisotropic colloidal platelets
Authors:
Carina Karner,
Christoph Dellago,
Emanuela Bianchi
Abstract:
Anisotropy at the level of the inter-particle interaction provides the particles with specific instructions for the self-assembly of target structures. The ability to synthesize non-spherical colloids, together with the possibility of controlling the particle bonding pattern via suitably placed interaction sites, is nowadays enlarging the playfield for materials design. We consider a model of anis…
▽ More
Anisotropy at the level of the inter-particle interaction provides the particles with specific instructions for the self-assembly of target structures. The ability to synthesize non-spherical colloids, together with the possibility of controlling the particle bonding pattern via suitably placed interaction sites, is nowadays enlarging the playfield for materials design. We consider a model of anisotropic colloidal platelets with regular rhombic shape and two attractive sites placed along adjacent edges and we run Monte Carlo simulations in two-dimensions to investigate the two-stage assembly of these units into clusters with well-defined symmetries and, subsequently, into extended lattices. Our focus is on how the site positioning and site-site attraction strength can be tuned to obtain micellar aggregates that are robust enough to successively undergo to a second-stage assembly from sparse clusters into a stable hexagonal lattice.
△ Less
Submitted 2 January, 2020;
originally announced January 2020.
-
How patchiness controls the properties of chain-like assemblies of colloidal platelets
Authors:
Carina Karner,
Christoph Dellago,
Emanuela Bianchi
Abstract:
Patchy colloidal platelets with convex, non-spherical shapes have been realized with different materials at length scales ranging from nanometers to microns. While the assembly of these hard shapes tends to maximize edge-to-edge contacts, as soon as a directional attraction is added -- by means of, e.g., specific ligands along the particle edges -- a competition between shape and bonding anisotrop…
▽ More
Patchy colloidal platelets with convex, non-spherical shapes have been realized with different materials at length scales ranging from nanometers to microns. While the assembly of these hard shapes tends to maximize edge-to-edge contacts, as soon as a directional attraction is added -- by means of, e.g., specific ligands along the particle edges -- a competition between shape and bonding anisotropy sets in, giving rise to a complex assembly scenario. We focus here on a two-dimensional system of patchy rhombi, i.e., colloidal platelets with a regular rhombic shape decorated with bonding sites along their perimeter. Specifically, we consider rhombi with two patches, placed on either opposite or adjacent edges. While for the first particle class only chains can form, for the latter we observe the emergence of either chains or loops, depending on the system parameters. According to the patch positioning -- classified in terms of different classes, topologies and distances from the edge center -- we are able to characterize the emerging chain-like assemblies in terms of length, packing abilities, flexibility properties and nematic ordering.
△ Less
Submitted 22 October, 2019; v1 submitted 15 October, 2019;
originally announced October 2019.
-
Design of patchy rhombi: from close-packed tilings to open lattices
Authors:
Carina Karner,
Emanuela Bianchi,
Christoph Dellago
Abstract:
In the realm of functional materials, the production of two-dimensional structures with tuneable porosity is of paramount relevance for many practical applications: surfaces with regular arrays of pores can be used for selective adsorption or immobilization of guest units that are complementary in shape and/or size to the pores, thus achieving, for instance, selective filtering or well-defined res…
▽ More
In the realm of functional materials, the production of two-dimensional structures with tuneable porosity is of paramount relevance for many practical applications: surfaces with regular arrays of pores can be used for selective adsorption or immobilization of guest units that are complementary in shape and/or size to the pores, thus achieving, for instance, selective filtering or well-defined responses to external stimuli. The principles that govern the formation of such structures are valid at both the molecular and the colloidal scale. Here we provide simple design directions to combine the anisotropic shape of the building units -- either molecules or colloids -- and selective directional bonding. Using extensive computer simulations we show that regular rhombic platelets decorated with attractive and repulsive interaction sites lead to specific tilings, going smoothly from close-packed arrangements to open lattices. The rationale behind the rich tiling scenario observed can be described in terms of steric incompatibilities, unsatisfied bonding geometries and interplays between local and long-range order.
△ Less
Submitted 3 July, 2019; v1 submitted 26 June, 2019;
originally announced June 2019.
