A study of why we need to reassess full reference image quality assessment with medical images

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.

A study on the adequacy of common IQA measures for medical images

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.

Distribution-informed and wavelength-flexible data-driven photoacoustic oximetry

Significance: Photoacoustic imaging (PAI) promises to measure spatially-resolved blood oxygen saturation, but suffers from a lack of accurate and robust spectral unmixing methods to deliver on this promise. Accurate blood oxygenation estimation could have important clinical applications, from cancer detection to quantifying inflammation. Aim: This study addresses the inflexibility of existing data-driven methods for estimating blood oxygenation in PAI by introducing a recurrent neural network architecture. Approach: We created 25 simulated training dataset variations to assess neural network performance. We used a long short-term memory network to implement a wavelength-flexible network architecture and proposed the Jensen-Shannon divergence to predict the most suitable training dataset. Results: The network architecture can handle arbitrary input wavelengths and outperforms linear unmixing and the previously proposed learned spectral decolouring method. Small changes in the training data significantly affect the accuracy of our method, but we find that the Jensen-Shannon divergence correlates with the estimation error and is thus suitable for predicting the most appropriate training datasets for any given application. Conclusions: A flexible data-driven network architecture combined with the Jensen-Shannon Divergence to predict the best training data set provides a promising direction that might enable robust data-driven photoacoustic oximetry for clinical use cases.

Moving beyond simulation: data-driven quantitative photoacoustic imaging using tissue-mimicking phantoms

Accurate measurement of optical absorption coefficients from photoacoustic imaging (PAI) data would enable direct mapping of molecular concentrations, providing vital clinical insight. The ill-posed nature of the problem of absorption coefficient recovery has prohibited PAI from achieving this goal in living systems due to the domain gap between simulation and experiment. To bridge this gap, we introduce a collection of experimentally well-characterised imaging phantoms and their digital twins. This first-of-a-kind phantom data set enables supervised training of a U-Net on experimental data for pixel-wise estimation of absorption coefficients. We show that training on simulated data results in artefacts and biases in the estimates, reinforcing the existence of a domain gap between simulation and experiment. Training on experimentally acquired data, however, yielded more accurate and robust estimates of optical absorption coefficients. We compare the results to fluence correction with a Monte Carlo model from reference optical properties of the materials, which yields a quantification error of approximately 20%. Application of the trained U-Nets to a blood flow phantom demonstrated spectral biases when training on simulated data, while application to a mouse model highlighted the ability of both learning-based approaches to recover the depth-dependent loss of signal intensity. We demonstrate that training on experimental phantoms can restore the correlation of signal amplitudes measured in depth. While the absolute quantification error remains high and further improvements are needed, our results highlight the promise of deep learning to advance quantitative PAI.

Unsupervised Domain Transfer with Conditional Invertible Neural Networks

Synthetic medical image generation has evolved as a key technique for neural network training and validation. A core challenge, however, remains in the domain gap between simulations and real data. While deep learning-based domain transfer using Cycle Generative Adversarial Networks and similar architectures has led to substantial progress in the field, there are use cases in which state-of-the-art approaches still fail to generate training images that produce convincing results on relevant downstream tasks. Here, we address this issue with a domain transfer approach based on conditional invertible neural networks (cINNs). As a particular advantage, our method inherently guarantees cycle consistency through its invertible architecture, and network training can efficiently be conducted with maximum likelihood training. To showcase our method's generic applicability, we apply it to two spectral imaging modalities at different scales, namely hyperspectral imaging (pixel-level) and photoacoustic tomography (image-level). According to comprehensive experiments, our method enables the generation of realistic spectral data and outperforms the state of the art on two downstream classification tasks (binary and multi-class). cINN-based domain transfer could thus evolve as an important method for realistic synthetic data generation in the field of spectral imaging and beyond.

Semantic segmentation of multispectral photoacoustic images using deep learning

Photoacoustic imaging has the potential to revolutionise healthcare due to the valuable information on tissue physiology that is contained in multispectral photoacoustic measurements. Clinical translation of the technology requires conversion of the high-dimensional acquired data into clinically relevant and interpretable information. In this work, we present a deep learning-based approach to semantic segmentation of multispectral photoacoustic images to facilitate the interpretability of recorded images. Manually annotated multispectral photoacoustic imaging data are used as gold standard reference annotations and enable the training of a deep learning-based segmentation algorithm in a supervised manner. Based on a validation study with experimentally acquired data of healthy human volunteers, we show that automatic tissue segmentation can be used to create powerful analyses and visualisations of multispectral photoacoustic images. Due to the intuitive representation of high-dimensional information, such a processing algorithm could be a valuable means to facilitate the clinical translation of photoacoustic imaging.

Data-driven generation of plausible tissue geometries for realistic photoacoustic image synthesis

Photoacoustic tomography (PAT) has the potential to recover morphological and functional tissue properties such as blood oxygenation with high spatial resolution and in an interventional setting. However, decades of research invested in solving the inverse problem of recovering clinically relevant tissue properties from spectral measurements have failed to produce solutions that can quantify tissue parameters robustly in a clinical setting. Previous attempts to address the limitations of model-based approaches with machine learning were hampered by the absence of labeled reference data needed for supervised algorithm training. While this bottleneck has been tackled by simulating training data, the domain gap between real and simulated images remains a huge unsolved challenge. As a first step to address this bottleneck, we propose a novel approach to PAT data simulation, which we refer to as "learning to simulate". Our approach involves subdividing the challenge of generating plausible simulations into two disjoint problems: (1) Probabilistic generation of realistic tissue morphology, represented by semantic segmentation maps and (2) pixel-wise assignment of corresponding optical and acoustic properties. In the present work, we focus on the first challenge. Specifically, we leverage the concept of Generative Adversarial Networks (GANs) trained on semantically annotated medical imaging data to generate plausible tissue geometries. According to an initial in silico feasibility study our approach is well-suited for contributing to realistic PAT image synthesis and could thus become a fundamental step for deep learning-based quantitative PAT.

Tattoo tomography: Freehand 3D photoacoustic image reconstruction with an optical pattern

Purpose: Photoacoustic tomography (PAT) is a novel imaging technique that can spatially resolve both morphological and functional tissue properties, such as the vessel topology and tissue oxygenation. While this capacity makes PAT a promising modality for the diagnosis, treatment and follow-up of various diseases, a current drawback is the limited field-of-view (FoV) provided by the conventionally applied 2D probes. Methods: In this paper, we present a novel approach to 3D reconstruction of PAT data (Tattoo tomography) that does not require an external tracking system and can smoothly be integrated into clinical workflows. It is based on an optical pattern placed on the region of interest prior to image acquisition. This pattern is designed in a way that a tomographic image of it enables the recovery of the probe pose relative to the coordinate system of the pattern. This allows the transformation of a sequence of acquired PA images into one common global coordinate system and thus the consistent 3D reconstruction of PAT imaging data. Results: An initial feasibility study conducted with experimental phantom data and in vivo forearm data indicates that the Tattoo approach is well-suited for 3D reconstruction of PAT data with high accuracy and precision. Conclusion: In contrast to previous approaches to 3D ultrasound (US) or PAT reconstruction, the Tattoo approach neither requires complex external hardware nor training data acquired for a specific application. It could thus become a valuable tool for clinical freehand PAT.

Invertible Neural Networks for Uncertainty Quantification in Photoacoustic Imaging

Multispectral photoacoustic imaging (PAI) is an emerging imaging modality which enables the recovery of functional tissue parameters such as blood oxygenation. However, the underlying inverse problems are potentially ill-posed, meaning that radically different tissue properties may - in theory - yield comparable measurements. In this work, we present a new approach for handling this specific type of uncertainty by leveraging the concept of conditional invertible neural networks (cINNs). Specifically, we propose going beyond commonly used point estimates for tissue oxygenation and converting single-pixel initial pressure spectra to the full posterior probability density. This way, the inherent ambiguity of a problem can be encoded with multiple modes in the output. Based on the presented architecture, we demonstrate two use cases which leverage this information to not only detect and quantify but also to compensate for uncertainties: (1) photoacoustic device design and (2) optimization of photoacoustic image acquisition. Our in silico studies demonstrate the potential of the proposed methodology to become an important building block for uncertainty-aware reconstruction of physiological parameters with PAI.

Deep learning for biomedical photoacoustic imaging: A review

Photoacoustic imaging (PAI) is a promising emerging imaging modality that enables spatially resolved imaging of optical tissue properties up to several centimeters deep in tissue, creating the potential for numerous exciting clinical applications. However, extraction of relevant tissue parameters from the raw data requires the solving of inverse image reconstruction problems, which have proven extremely difficult to solve. The application of deep learning methods has recently exploded in popularity, leading to impressive successes in the context of medical imaging and also finding first use in the field of PAI. Deep learning methods possess unique advantages that can facilitate the clinical translation of PAI, such as extremely fast computation times and the fact that they can be adapted to any given problem. In this review, we examine the current state of the art regarding deep learning in PAI and identify potential directions of research that will help to reach the goal of clinical applicability