Bidirectional meta-Kronecker factored optimizer and Hausdorff distance loss for few-shot medical image segmentation.

To increase the accuracy of medical image analysis using supervised learning-based AI technology, a large amount of accurately labeled training data is required. However, the supervised learning approach may not be applicable to real-world medical imaging due to the lack of labeled data, the privacy of patients, and the cost of specialized knowledge. To handle these issues, we utilized Kronecker-factored decomposition, which enhances both computational efficiency and stability of the learning process. We combined this approach with a model-agnostic meta-learning framework for the parameter optimization. Based on this method, we present a bidirectional meta-Kronecker factored optimizer (BM-KFO) framework to quickly optimize semantic segmentation tasks using just a few magnetic resonance imaging (MRI) images as input. This model-agnostic approach can be implemented without altering network components and is capable of learning the learning process and meta-initial points while training on previously unseen data. We also incorporated a combination of average Hausdorff distance loss (AHD-loss) and cross-entropy loss into our objective function to specifically target the morphology of organs or lesions in medical images. Through evaluation of the proposed method on the abdominal MRI dataset, we obtained an average performance of 78.07% in setting 1 and 79.85% in setting 2. Our experiments demonstrate that BM-KFO with AHD-loss is suitable for general medical image segmentation applications and achieves superior performance compared to the baseline method in few-shot learning tasks. In order to replicate the proposed method, we have shared our code on GitHub. The corresponding URL can be found: https://github.com/YeongjoonKim/BMKFO.git .

Uniformity Attentive Learning-Based Siamese Network for Person Re-Identification.

Person re-identification (Re-ID) has a problem that makes learning difficult such as misalignment and occlusion. To solve these problems, it is important to focus on robust features in intra-class variation. Existing attention-based Re-ID methods focus only on common features without considering distinctive features. In this paper, we present a novel attentive learning-based Siamese network for person Re-ID. Unlike existing methods, we designed an attention module and attention loss using the properties of the Siamese network to concentrate attention on common and distinctive features. The attention module consists of channel attention to select important channels and encoder-decoder attention to observe the whole body shape. We modified the triplet loss into an attention loss, called uniformity loss. The uniformity loss generates a unique attention map, which focuses on both common and discriminative features. Extensive experiments show that the proposed network compares favorably to the state-of-the-art methods on three large-scale benchmarks including Market-1501, CUHK03 and DukeMTMC-ReID datasets.

Energy Calibration of a CdTe Photon Counting Spectral Detector with Consideration of its Non-Convergent Behavior.

Fast and accurate energy calibration of photon counting spectral detectors (PCSDs) is essential for their biomedical applications to identify and characterize bio-components or contrast agents in tissues. Using the x-ray tube voltage as a reference for energy calibration is known to be an efficient method, but there has been no consideration in the energy calibration of non-convergent behavior of PCSDs. We observed that a single pixel mode (SPM) CdTe PCSD based on Medipix-2 shows some non-convergent behaviors in turning off the detector elements when a high enough threshold is applied to the comparator that produces a binary photon count pulse. More specifically, the detector elements are supposed to stop producing photon count pulses once the threshold reaches a point of the highest photon energy determined by the tube voltage. However, as the x-ray exposure time increases, the threshold giving 50% of off pixels also increases without converging to a point. We established a method to take account of the non-convergent behavior in the energy calibration. With the threshold-to-photon energy mapping function established by the proposed method, we could better identify iodine component in a phantom consisting of iodine and other components.

Tissue Cancellation in Dual Energy Mammography Using a Calibration Phantom Customized for Direct Mapping.

An easily implementable tissue cancellation method for dual energy mammography is proposed to reduce anatomical noise and enhance lesion visibility. For dual energy calibration, the images of an imaging object are directly mapped onto the images of a customized calibration phantom. Each pixel pair of the low and high energy images of the imaging object was compared to pixel pairs of the low and high energy images of the calibration phantom. The correspondence was measured by absolute difference between the pixel values of imaged object and those of the calibration phantom. Then the closest pixel pair of the calibration phantom images is marked and selected. After the calibration using direct mapping, the regions with lesion yielded different thickness from the background tissues. Taking advantage of the different thickness, the visibility of cancerous lesions was enhanced with increased contrast-to-noise ratio, depending on the size of lesion and breast thickness. However, some tissues near the edge of imaged object still remained after tissue cancellation. These remaining residuals seem to occur due to the heel effect, scattering, nonparallel X-ray beam geometry and Poisson distribution of photons. To improve its performance further, scattering and the heel effect should be compensated.

A unified statistical framework for material decomposition using multienergy photon counting x-ray detectors.

PURPOSE: Material decomposition using multienergy photon counting x-ray detectors (PCXD) has been an active research area over the past few years. Even with some success, the problem of optimal energy selection and three material decomposition including malignant tissue is still on going research topic, and more systematic studies are required. This paper aims to address this in a unified statistical framework in a mammographic environment. METHODS: A unified statistical framework for energy level optimization and decomposition of three materials is proposed. In particular, an energy level optimization algorithm is derived using the theory of the minimum variance unbiased estimator, and an iterative algorithm is proposed for material composition as well as system parameter estimation under the unified statistical estimation framework. To verify the performance of the proposed algorithm, the authors performed simulation studies as well as real experiments using physical breast phantom and ex vivo breast specimen. Quantitative comparisons using various performance measures were conducted, and qualitative performance evaluations for ex vivo breast specimen were also performed by comparing the ground-truth malignant tissue areas identified by radiologists. RESULTS: Both simulation and real experiments confirmed that the optimized energy bins by the proposed method allow better material decomposition quality. Moreover, for the specimen thickness estimation errors up to 2 mm, the proposed method provides good reconstruction results in both simulation and real ex vivo breast phantom experiments compared to existing methods. CONCLUSIONS: The proposed statistical framework of PCXD has been successfully applied for the energy optimization and decomposition of three material in a mammographic environment. Experimental results using the physical breast phantom and ex vivo specimen support the practicality of the proposed algorithm.

Automatic registration between 3D intra-operative ultrasound and pre-operative CT images of the liver based on robust edge matching.

The registration of a three-dimensional (3D) ultrasound (US) image with a computed tomography (CT) or magnetic resonance image is beneficial in various clinical applications such as diagnosis and image-guided intervention of the liver. However, conventional methods usually require a time-consuming and inconvenient manual process for pre-alignment, and the success of this process strongly depends on the proper selection of initial transformation parameters. In this paper, we present an automatic feature-based affine registration procedure of 3D intra-operative US and pre-operative CT images of the liver. In the registration procedure, we first segment vessel lumens and the liver surface from a 3D B-mode US image. We then automatically estimate an initial registration transformation by using the proposed edge matching algorithm. The algorithm finds the most likely correspondences between the vessel centerlines of both images in a non-iterative manner based on a modified Viterbi algorithm. Finally, the registration is iteratively refined on the basis of the global affine transformation by jointly using the vessel and liver surface information. The proposed registration algorithm is validated on synthesized datasets and 20 clinical datasets, through both qualitative and quantitative evaluations. Experimental results show that automatic registration can be successfully achieved between 3D B-mode US and CT images even with a large initial misalignment.

Three-dimensional blood vessel quantification via centerline deformation.

It is clinically important to quantify the geometric parameters of an abnormal vessel, as this information can aid radiologists in choosing appropriate treatments or apparatuses. Centerline and cross-sectional diameters are commonly used to characterize the morphology of vessel in various clinical applications. Due to the existence of stenosis or aneurysm, the associated vessel centerline is unable to truly portray the original, healthy vessel shape and may result in inaccurate quantitative measurement. To remedy such a problem, a novel method using an active tube model is proposed. In the method, a smoothened centerline is determined as the axis of a deformable tube model that is registered onto the vessel lumen. Three types of regions, normal, stenotic, and aneurysmal regions, are defined to classify the vessel segment under-analyzed by use of the algorithm of a cross-sectional-based distance field. The registration process used on the tube model is governed by different region-adaptive energy functionals associated with the classified vessel regions. The proposed algorithm is validated on the 3-D computer-generated phantoms and 3-D rotational digital subtraction angiography (DSA) datasets. Experimental results show that the deformed centerline provides better vessel quantification results compared with the original centerline. It is also shown that the registered model is useful for measuring the volume of aneurysmal regions.

A new path planning algorithm for maximizing visibility in computed tomography colonography.

In virtual colonoscopy, minimizing the blind areas is important for accurate diagnosis of colonic polyps. Although useful for describing the shape of an object, the centerline is not always the optimal camera path for observing the object. Hence, conventional methods in which the centerline is directly used as a path produce considerable blind areas, especially in areas of high curvature. Our proposed algorithm first approximates the surface of the object by estimating the overall shape and cross-sectional thicknesses. View positions and their corresponding view directions are then jointly determined to enable us to maximally observe the approximated surface. Moreover, by adopting bidirectional navigations, we may reduce the blind area blocked by haustral folds. For comfortable navigation, we carefully smoothen the obtained path and minimize the amount of rotation between consecutive rendered images. For the evaluation, we quantified the overall observable area on the basis of the temporal visibility that reflects the minimum interpretation time of a human observer. The experimental results show that our algorithm improves visibility coverage and also significantly reduces the number of blind areas that have a clinically meaningful size. A sequence of rendered images shows that our algorithm can provide a sequence of centered and comfortable views of colonography.