Lifelong learning on Whole Slide Images (WSIs) aims to train or fine-tune a unified model sequentially on cancer-related tasks, reducing the resources and effort required for data transfer and processing, especially given the gigabyte-scale size of WSIs. In this paper, we introduce MergeSlide, a simple yet effective framework that treats lifelong learning as a model-merging problem by leveraging a vision–language pathology foundation model. When a new task arrives, it is ❶ defined with class-aware prompts, ❷ fine-tuned for a few epochs in a classifier-free manner, and ❸ merged into a unified model using an orthogonal continual-merging strategy that preserves performance and mitigates catastrophic forgetting. For inference under the class-incremental learning (CLASS-IL) setting, where task identity is unknown, we introduce Task-to-Class Prompt-aligned (TCP) inference. Specifically, TCP first identifies the most relevant task using task-level prompts and then applies the corresponding class-aware prompts to generate predictions. To evaluate MergeSlide, we conduct experiments on a stream of six TCGA datasets. The results show that MergeSlide outperforms both rehearsal-based continual learning and vision-language zero-shot baselines. Code and data are available at https://github.com/caodoanh2001/MergeSlide.

Inverse problems span across diverse fields. In medical contexts, computed tomography (CT) plays a crucial role in reconstructing a patient's internal structure, presenting challenges due to artifacts caused by inherently ill-posed inverse problems. Previous research advanced image quality via post-processing and deep unrolling algorithms but faces challenges, such as extended convergence times with ultra-sparse data. Despite enhancements, resulting images often show significant artifacts, limiting their effectiveness for real-world diagnostic applications. We aim to explore deep second-order unrolling algorithms for solving imaging inverse problems, emphasizing their faster convergence and lower time complexity compared to common first-order methods like gradient descent. In this paper, we introduce QN-Mixer, an algorithm based on the quasi-Newton approach. We use learned parameters through the BFGS algorithm and introduce Incept-Mixer, an efficient neural architecture that serves as a non-local regularization term, capturing long-range dependencies within images. To address the computational demands typically associated with quasi-Newton algorithms that require full Hessian matrix computations, we present a memory-efficient alternative. Our approach intelligently downsamples gradient information, significantly reducing computational requirements while maintaining performance. The approach is validated through experiments on the sparse-view CT problem, involving various datasets and scanning protocols, and is compared with post-processing and deep unrolling state-of-the-art approaches. Our method outperforms existing approaches and achieves state-of-the-art performance in terms of SSIM and PSNR, all while reducing the number of unrolling iterations required.

Fingerprint classification is a successful technique to reduce the search space of a large fingerprint database in identification systems. It is based on determining the class of a query fingerprint. In this work, we investigate the use of Conic Radon Transform (CRT) as a feature extractor and a deep learning technique in order to solve fingerprint classification tasks. The proposed approach is based on the CRT. The latter represents an extension of classical Radon Transform (RT) to integrate an image function f(x, y) over conic sections. The Radon technique enables the extraction of fingerprint's global characteristics which are invariant to geometrical transformations such as translation and rotation. We perform CRT on a source image, then we combine the Radon result with an original image and use them as an input for convolutional neural networks (CNN). In our experiments, our approach has yielded better accuracy than state-of-the-art methods based on explicit feature extraction, even others CNN methods based on original images.

This paper presents a Generalized Multi Scaled Radon Transform (GMSRT) which can detect parametric curves and objects of any position, orientation and scale. This new transform extends the Generalized Multi Directional Radon Transform (GMDRT) in order to outperform its recognition rates. Despite the great success of GMDRT for the geometric shape detection, it still limited in dealing with the scale variation issue. Our proposed approach combines the GMDRT and Wavelet Transform (WT) to recognize shapes with different scales. We have observed a clear improvement in terms of classification accuracy compared to the GMDRT. Experiments of the proposed approach is done on the MPEG7 dataset. Comparison with some previous approaches demonstrates the efficiency of the proposed approach in detecting complex objects, even under geometric transformations.

Bimodal medical imaging, which combines two non-invasive techniques to explore two different aspects (e.g. function and morphology) of an organ, has emerged as a major clinical imaging modality nowadays. The gained information as compared to single modality imaging is overwhelmingly useful for diagnostics and therapy planning. Up to now all bimodal techniques make use exclusively of primary radiation. In this work, we describe a new bimodal imaging concept based solely on the exploitation of scattered radiation. Following a presentation of its theoretical basis, we discuss the modellings of measurements and inversion formulae used for image reconstruction. Thus an attenuation and electron density correction algorithm inspired of the Iterative Pre-Correction algorithm (IPC) is proposed. Simulation results demonstrate convincingly the feasibility and efficiency of this new bimodal imaging modality.

The rigourous mathematical model for radar emission and scattering is given by Maxwell's equations of electromagnetism. But due to the complexity of deriving a reconstruction formula based directly on these equations, other solutions are prefered to solve the mathematical problems in Synthetic Aperture Radar (SAR). Thus the modelling of SAR by a spherical Radon transform (SRT) appears as an interesting approach for the inversion of such data and could be used in some cases like high squint and ultra-wideband SAR. In this paper, a new algorithm is proposed for the stable inversion of measurement (SRT) and simulation results demonstrate the theoretical feasibility and viability of the algorithm.

The Radon transform (RT) on straight lines deals as mathematical foundation for conventional tomographic imaging (e.g. X-ray scanner, Single Photon Emission Computed Tomography (SPECT), Positron Emission Tomography (PET)) which use only one physical phenomenon, i.e. either transmission or emission of radiation. An imaging concept exploiting jointly two phenomena leads to a Radon transform defined on a geometrical support different from straight lines. In this paper, we propose a new two-dimensional X-ray imaging based on the coupling between transmission and reflection. Its modeling leads to a Radon transform defined on a pair of half-lines forming a vertical letter V, called V-line RT (V-RT). Moreover we establish the analytic inverse formula of this new transform, which forms the mathematical basis for image reconstruction. Through simulations, image formation and reconstruction results show the feasibility of this new imaging. The main advantage is the use of an one-dimensional detector which does not rotate for two-dimensional image reconstruction.

Conventional tomographies (e.g. X-ray scanner, Positron Emission Tomography), functioning on primary radiation are modeled by the Radon transform (RT) which is an integral transform defined on straight lines. In this paper, we consider a new RT defined on a pair of half-lines forming a letter V, which arises from modeling two-dimensional emission imaging process by Compton scattered gamma rays. We give its definition and establish its analytic inverse to support the feasibility of the reconstruction of a two-dimensional (2D) image from scattered radiation measured on a one-dimensional collimated camera. Moreover, a filtered back-projection inversion method is also constructed. Its main advantages are algorithmic efficiency and computational rapidity. We present numerical simulations to illustrate the working. To sum up, the V-line RT leads not only to a new imaging principle, but also to a new concept of detector with high energetic resolution capable to collect the scattered radiation.

We propose a new scheme for imaging radiation emitting objects by measuring its non collimated scattered radiation at various energies. A few years ago we have put forward the idea of scatter imaging and show how it works with a collimated detector. To increase drastically the sensitivity of this modality as well as its field of view and resolution we propose that data acquisition should be performed without collimation as it is done in Compton cameras. We discuss image formation by scattered radiation in this context by computing and comparing the related point spread functions (PSF) and comment on their properties. We also present numerical simulations to support the attractiveness of this modality.

This paper presents a new dual-modality imaging principle using two recent results. The first one demonstrates the feasibility of reconstruction of a radioactive distribution from its Compton-scattered radiation. This may be regarded as a novel gamma-ray emission imaging principle. The second one shows the possibility to reconstruct the electronic density of a medium and its attenuation map from other Compton-scattered radiation emitted by an external gamma source and scattered in the medium. The required data for the two reconstructions are easily acquired from an energy and space measuring gamma camera under the form of scattered distribution images classified by their Compton-scattering angle. The usual motion of camera is no longer necessary and so all images needed for a three-dimensional reconstruction are recorded simultaneously. For non-immobile object, this is a decisive advantage

Spherical maps occur in a range of applications for instance in geophysics or in astrophysics with the study of the cosmic microwave background (CMB) radiation field, where observations are over the whole sky. Analyzing these images requires specific tools. This paper describes a new multiscale decomposition for data on the sphere, namely the curvelet transform on the sphere. The curvelet transform, in its first step, requires the use of an isotropic wavelet transform. Therefore, our new curvelet transform also includes a new wavelet transform on the sphere which has properties similar to those of the a trous isotropic wavelet transform.

This paper describes a new deconvolution algorithm, based on both the wavelet transform and the curvelet transform. It extends previous results which were obtained for the denoising problem. Using these two different transformations in the same algorithm allows us to optimally detect in the same time isotropic features, well represented by the wavelet transform, and edges better represented by the curvelet transform. Adding a TV penalization term avoids the presence of oscillatory patterns around the edges which may appear when using multiscale methods. We illustrate the results with simulations.

The solution to the scatter problem remains a major technical challenge in gamma ray imaging. This paper presents a modelization of apparent images formed by scattered photons at various energies. It may be used either to generate a series of realistic multi-energy images or be included in an image enhancement procedure. The application to SPECT is discussed.

We are interested in the problem of restoring scintigraphic images acquired by a gamma detector in nuclear medicine. The aim is to improve the detectability of possible heterogeneous areas in different organs. We propose to solve the problem in the Bayesian framework with the maximum a posteriori (MAP) principle. The prior information was modeled by a Markov random field (MRF). The optimization is based on two kinds of methods: the stochastic algorithm of simulated annealing with a Gibbs sampler, and the deterministic algorithm of graduated non-convexity (GNC). We compared the results to the images restored by the Metz filter, more classical in this field. We applied these methods to the restoration of cold or warm nodules in the thyroid gland. We noticed the superiority of the proposed methods in terms of contrast around the nodules and uniformity in the images.

Scintigraphic imagery is widely used in nuclear medicine and in industrial resting. However, the image quality is vary poor due to several degradations: Poisson noise, scattering of gamma photons, non-stationary impulse response of the gamma detector. The restoration of scintigraphic images is typically an ill-posed inverse problem. In this paper, we propose a restoration method based on the Bayes-Markov approach. The regularization of such a problem is carried out by a Markovian prior. The discontinuity recovery, and the restoration of the homogenous areas are improved thanks to the Markov random field (MRF) with an implicit line process. The performance of this approach is shown through the quality measures in terms of contrast around the edges and uniformity in the images, in comparison with two other existing methods.

Microwave imaging is of great interest in medical applications owing to its high sensitivity with respect to dielectric properties. It allows detection of very small inhomogeneities. The image reconstruction employing the microwave inverse scattering consists of reconstructing the image of an object from the scattered field measured behind the object. This reconstruction runs up against the nonuniqueness of the solution of the inverse scattering problem. The authors propose to solve the ill-posed inverse problem by a statistical regularization method based on the Bayesian maximum a posteriori estimation where the principle of maximum entropy is used for assigning the a priori laws. The results obtained demonstrate the power and potential of this method in image reconstruction.