Quality control (QC) of MR images is essential to ensure that downstream analyses such as segmentation can be performed successfully. Currently, QC is predominantly performed visually and subjectively, at significant time and operator cost. We aim to
automate the process using a probabilistic network that estimates segmentation uncertainty through a heteroscedastic noise model, providing a measure of task-specific quality. By augmenting training images with k-space artefacts, we propose a novel CNN architecture to decouple sources of uncertainty related to the task and different k-space artefacts in a self-supervised manner. This enables the prediction of separate uncertainties for different types of data degradation. While the uncertainty predictions reflect the presence and severity of artefacts, the network provides more robust and generalisable segmentation predictions given the quality of the data. We show that models trained with artefact augmentation provide informative measures of uncertainty on both simulated artefacts and problematic real-world images identified by human raters, both qualitatively and quantitatively in the form of error bars on volume measurements. Relating artefact uncertainty to segmentation Dice scores, we observe that our uncertainty predictions provide a better estimate of MRI quality from the point of view of the task (gray matter segmentation) compared to commonly used metrics of quality including signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR), hence providing a real-time quality metric indicative of segmentation quality.
Quality control (QC) in medical image analysis is time-consuming and laborious, leading to increased interest in automated methods. However, what is deemed suitable quality for algorithmic processing may be different from human-perceived measures of
visual quality. In this work, we pose MR image quality assessment from an image reconstruction perspective. We train Bayesian CNNs using a heteroscedastic uncertainty model to recover clean images from noisy data, providing measures of uncertainty over the predictions. This framework enables us to divide data corruption into learnable and non-learnable components and leads us to interpret the predictive uncertainty as an estimation of the achievable recovery of an image. Thus, we argue that quality control for visual assessment cannot be equated to quality control for algorithmic processing. We validate this statement in a multi-task experiment combining artefact recovery with uncertainty prediction and grey matter segmentation. Recognising this distinction between visual and algorithmic quality has the impact that, depending on the downstream task, less data can be excluded based on ``visual quality reasons alone.
Multimode fibres are becoming increasingly attractive in optical endoscopy as they promise to enable unparalleled miniaturisation, spatial resolution and cost as compared to conventional fibre bundle-based counterpart. However, achieving high-speed i
maging through a multimode fibre (MMF) based on wavefront shaping has been challenging due to the use of liquid crystal spatial light modulators with low frame rates. In this work, we report the development of a video-rate dual-modal forward-viewing photoacoustic (PA) and fluorescence endo-microscopy probe based on a MMF and a high-speed digital micromirror device (DMD). Light transmission characteristics through the fibre were characterised with a real-valued intensity transmission matrix algorithm, and subsequently, optimal binary patterns were calculated to focus light through the fibre with wavefront shaping. Raster-scanning of a tightly focused beam (1.5 {mu}m diameter) at the distal end of the fibre was performed for imaging. With the DMD running at 10 kHz, the PA imaging speed and spatial resolution of were controlled by varying the scanning step size, ranging from 1 to 25 frames per second (fps) and from 1.7 to 3 {mu}m, respectively, over a field-of-view of 50 {mu}m x 50 {mu}m. High-resolution PA images of carbon fibres, and mouse red blood cells were acquired through a MMF with high image fidelity at unprecedented speed with MMF-based PA endoscope. The capability of dual-modal PA and fluorescence imaging was demonstrated by imaging phantoms comparing carbon fibres and fluorescent microspheres. We anticipate that with further miniaturisation of the ultrasound detector, this probe could be integrated into a medical needle to guide minimally invasive procedures in several clinical contexts including tumour biopsy and nerve blocks.
In fetal Magnetic Resonance Imaging, Super Resolution Reconstruction (SRR) algorithms are becoming popular tools to obtain high-resolution 3D volume reconstructions from low-resolution stacks of 2D slices, acquired at different orientations. To be ef
fective, these algorithms often require accurate segmentation of the region of interest, such as the fetal brain in suspected pathological cases. In the case of Spina Bifida, Ebner, Wang et al. (NeuroImage, 2020) combined their SRR algorithm with a 2-step segmentation pipeline (2D localisation followed by a 2D segmentation network). However, if the localisation step fails, the second network is not able to recover a correct brain mask, thus requiring manual corrections for an effective SRR. In this work, we aim at improving the fetal brain segmentation for SRR in Spina Bifida. We hypothesise that a well-trained single-step UNet can achieve accurate performance, avoiding the need of a 2-step approach. We propose a new tool for fetal brain segmentation called MONAIfbs, which takes advantage of the Medical Open Network for Artificial Intelligence (MONAI) framework. Our network is based on the dynamic UNet (dynUNet), an adaptation of the nnU-Net framework. When compared to the original 2-step approach proposed in Ebner-Wang, and the same Ebner-Wang approach retrained with the expanded dataset available for this work, the dynUNet showed to achieve higher performance using a single step only. It also showed to reduce the number of outliers, as only 28 stacks obtained Dice score less than 0.9, compared to 68 for Ebner-Wang and 53 Ebner-Wang expanded. The proposed dynUNet model thus provides an improvement of the state-of-the-art fetal brain segmentation techniques, reducing the need for manual correction in automated SRR pipelines. Our code and our trained model are made publicly available at https://github.com/gift-surg/MONAIfbs.
Training a deep neural network is an optimization problem with four main ingredients: the design of the deep neural network, the per-sample loss function, the population loss function, and the optimizer. However, methods developed to compete in recen
t BraTS challenges tend to focus only on the design of deep neural network architectures, while paying less attention to the three other aspects. In this paper, we experimented with adopting the opposite approach. We stuck to a generic and state-of-the-art 3D U-Net architecture and experimented with a non-standard per-sample loss function, the generalized Wasserstein Dice loss, a non-standard population loss function, corresponding to distributionally robust optimization, and a non-standard optimizer, Ranger. Those variations were selected specifically for the problem of multi-class brain tumor segmentation. The generalized Wasserstein Dice loss is a per-sample loss function that allows taking advantage of the hierarchical structure of the tumor regions labeled in BraTS. Distributionally robust optimization is a generalization of empirical risk minimization that accounts for the presence of underrepresented subdomains in the training dataset. Ranger is a generalization of the widely used Adam optimizer that is more stable with small batch size and noisy labels. We found that each of those variations of the optimization of deep neural networks for brain tumor segmentation leads to improvements in terms of Dice scores and Hausdorff distances. With an ensemble of three deep neural networks trained with various optimization procedures, we achieved promising results on the validation dataset of the BraTS 2020 challenge. Our ensemble ranked fourth out of the 693 registered teams for the segmentation task of the BraTS 2020 challenge.
Limiting failures of machine learning systems is vital for safety-critical applications. In order to improve the robustness of machine learning systems, Distributionally Robust Optimization (DRO) has been proposed as a generalization of Empirical Ris
k Minimization (ERM)aiming at addressing this need. However, its use in deep learning has been severely restricted due to the relative inefficiency of the optimizers available for DRO in comparison to the wide-spread variants of Stochastic Gradient Descent (SGD) optimizers for ERM. We propose SGD with hardness weighted sampling, a principled and efficient optimization method for DRO in machine learning that is particularly suited in the context of deep learning. Similar to a hard example mining strategy in essence and in practice, the proposed algorithm is straightforward to implement and computationally as efficient as SGD-based optimizers used for deep learning, requiring minimal overhead computation. In contrast to typical ad hoc hard mining approaches, and exploiting recent theoretical results in deep learning optimization, we prove the convergence of our DRO algorithm for over-parameterized deep learning networks with ReLU activation and finite number of layers and parameters. Our experiments on brain tumor segmentation in MRI demonstrate the feasibility and the usefulness of our approach. Using our hardness weighted sampling leads to a decrease of 2% of the interquartile range of the Dice scores for the enhanced tumor and the tumor core regions. The code for the proposed hard weighted sampler will be made publicly available.
Reconstruction of PET images is an ill-posed inverse problem and often requires iterative algorithms to achieve good image quality for reliable clinical use in practice, at huge computational costs. In this paper, we consider the PET reconstruction a
dense prediction problem where the large scale contextual information is essential, and propose a novel architecture of multi-scale fully convolutional neural networks (msfCNN) for fast PET image reconstruction. The proposed msfCNN gains large receptive fields with both memory and computational efficiency, by using a downscaling-upscaling structure and dilated convolutions. Instead of pooling and deconvolution, we propose to use the periodic shuffling operation from sub-pixel convolution and its inverse to scale the size of feature maps without losing resolution. Residual connections were added to improve training. We trained the proposed msfCNN model with simulated data, and applied it to clinical PET data acquired on a Siemens mMR scanner. The results from real oncological and neurodegenerative cases show that the proposed msfCNN-based reconstruction outperforms the iterative approaches in terms of computational time while achieving comparable image quality for quantification. The proposed msfCNN model can be applied to other dense prediction tasks, and fast msfCNN-based PET reconstruction could facilitate the potential use of molecular imaging in interventional/surgical procedures, where cancer surgery can particularly benefit.
