Instrumentation Institute For Molecular Image

Castillo D., Rodríguez-Álvarez M.J., Samaniego R., Lakshminarayanan V.

According to the World Health Organization (WHO), peripheral and central neurological disorders affect approximately one billion people worldwide. Ischemic stroke and Alzheimer’s Disease and other dementias are the second and fifth leading causes of death, respectively. In this context, detecting and classifying brain lesions constitute a critical area of research in medical image processing, significantly impacting clinical practice. Traditional lesion detection, segmentation, and feature extraction methods are time-consuming and observer-dependent. In this sense, research in the machine and deep learning methods applied to medical image processing constitute one of the crucial tools for automatically learning hierarchical features to get better accuracy, quick diagnosis, treatment, and prognosis of diseases. This project aims to develop and implement deep learning models for detecting and classifying small brain White Matter hyperintensities (WMH) lesions in magnetic resonance images (MRI), specifically lesions concerning ischemic and demyelination diseases. The methods applied were the UNet and Segmenting Anything model (SAM) for segmentation, while YOLOV8 and Detectron2 (based on MaskRCNN) were also applied to detect and classify the lesions. Experimental results show a Dice coefficient (DSC) of 0.94, 0.50, 0.241, and 0.88 for segmentation of WMH lesions using the UNet, SAM, YOLOv8, and Detectron2, respectively. The Detectron2 model demonstrated an accuracy of 0.94 in detecting and 0.98 in classifying lesions, including small lesions where other models often fail. The methods developed give an outline for the detection, segmentation, and classification of small and irregular morphology brain lesions and could significantly aid clinical diagnostics, providing reliable support for physicians and improving patient outcomes.

Ramos M.R., López J.G., Seimetz M., Morales J.J., Muñoz C.T., Ramos M.D.

The development of advanced detection systems for charged particles in laser-based accelerators and the need for precise time of flight measurements have led to the creation of detectors using ultra-thin plastic scintillators, indicating their use as transmission detectors with low energy loss and minimal dispersion for protons around a few MeV. This study introduces a new detection system designed by the Institute for Instrumentation in Molecular Imaging for time of flight and timing applications at the National Accelerator Center in Seville. The system includes an ultra-thin EJ-214 plastic scintillator coupled with a photomultiplier tube and shielded by aluminized mylar sheets. The prototype installation as an external trigger system at the ion beam nuclear microprobe of the aforementioned facility, along with its temporal performance and ion transmission, was thoroughly characterized. Additionally, the scintillator thickness and uniformity were analyzed using Rutherford backscattering spectrometry. Results showed that the experimental thickness of the EJ-214 sheet differs by approximately 46% from the supplier specifications. The detector response to MeV protons demonstrates a strong dependence on the impact position but remains mostly linear with the applied working bias. Finally, single ion detection was successfully achieved, demonstrating the applicability of this new system as a diagnostic tool.

Grattarola A., Mora M.C., Cerdá-Boluda J., Ortiz J.V.

Every year, thousands of people undergo amputations due to trauma or medical conditions. The loss of an upper limb, in particular, has profound physical and psychological consequences for patients. One potential solution is the use of externally powered prostheses equipped with motorized artificial hands. However, these commercially available prosthetic hands are prohibitively expensive for most users. In recent years, advancements in 3D printing and sensor technologies have enabled the design and production of low-cost, externally powered prostheses. This paper presents a pattern-recognition-based human–prosthesis interface that utilizes surface electromyography (sEMG) signals, captured by an affordable device, the Myo armband. A Support Vector Machine (SVM) algorithm, optimized using Bayesian techniques, is trained to classify the user’s intended grasp from among nine common grasping postures essential for daily life activities and functional prosthetic performance. The proposal is viable for real-time implementations on low-cost platforms with 85% accuracy in grasping posture recognition.

Nguyen G., Sáinz-Pardo Díaz J., Calatrava A., Berberi L., Lytvyn O., Kozlov V., Tran V., Moltó G., López García Á.

Abstract Machine learning is one of the most widely used technologies in the field of Artificial Intelligence. As machine learning applications become increasingly ubiquitous, concerns about data privacy and security have also grown. The work in this paper presents a broad theoretical landscape concerning the evolution of machine learning and deep learning from centralized to distributed learning, first in relation to privacy-preserving machine learning and secondly in the area of privacy-enhancing technologies. It provides a comprehensive landscape of the synergy between distributed machine learning and privacy-enhancing technologies, with federated learning being one of the most prominent architectures. Various distributed learning approaches to privacy-aware techniques are structured in a review, followed by an in-depth description of relevant frameworks and libraries, more particularly in the context of federated learning. The paper also highlights the need for data protection and privacy addressed from different approaches, key findings in the field concerning AI applications, and advances in the development of related tools and techniques.

Céspedes Sisniega J., Rodríguez V., Moltó G., López García Á.

As machine learning models are increasingly deployed in production, robust monitoring and detection of concept and covariate drift become critical. This paper addresses the gap in the widespread adoption of drift detection techniques by proposing a serverless-based approach for batch covariate drift detection in ML systems. Leveraging the open-source OSCAR framework and the open-source Frouros drift detection library, we develop a set of services that enable parallel execution of two key components: the ML inference pipeline and the batch covariate drift detection pipeline. To this end, our proposal takes advantage of the elasticity and efficiency of serverless computing for ML pipelines, including scalability, cost-effectiveness, and seamless integration with existing infrastructure. We evaluate this approach through an edge ML use case, showcasing its operation on a simulated batch covariate drift scenario. Our research highlights the importance of integrating drift detection as a fundamental requirement in developing robust and trustworthy AI systems and encourages the adoption of these techniques in ML deployment pipelines. In this way, organizations can proactively identify and mitigate the adverse effects of covariate drift while capitalizing on the benefits offered by serverless computing.

Jiménez-Gaona Y., Rodríguez-Alvarez M.J., Escudero L., Sandoval C., Lakshminarayanan V.

INTRODUCTION: Ultrasound in conjunction with mammography imaging, plays a vital role in the early detection and diagnosis of breast cancer. However, speckle noise affects medical ultrasound images and degrades visual radiological interpretation. Speckle carries information about the interactions of the ultrasound pulse with the tissue microstructure, which generally causes several difficulties in identifying malignant and benign regions. The application of deep learning in image denoising has gained more attention in recent years. OBJECTIVES: The main objective of this work is to reduce speckle noise while preserving features and details in breast ultrasound images using GAN models. METHODS: We proposed two GANs models (Conditional GAN and Wasserstein GAN) for speckle-denoising public breast ultrasound databases: BUSI, DATASET A, AND UDIAT (DATASET B). The Conditional GAN model was trained using the Unet architecture, and the WGAN model was trained using the Resnet architecture. The image quality results in both algorithms were measured by Peak Signal to Noise Ratio (PSNR, 35–40 dB) and Structural Similarity Index (SSIM, 0.90–0.95) standard values. RESULTS: The experimental analysis clearly shows that the Conditional GAN model achieves better breast ultrasound despeckling performance over the datasets in terms of PSNR = 38.18 dB and SSIM = 0.96 with respect to the WGAN model (PSNR = 33.0068 dB and SSIM = 0.91) on the small ultrasound training datasets. CONCLUSIONS: The observed performance differences between CGAN and WGAN will help to better implement new tasks in a computer-aided detection/diagnosis (CAD) system. In future work, these data can be used as CAD input training for image classification, reducing overfitting and improving the performance and accuracy of deep convolutional algorithms.

Sanz-Sanchez A., García F.B., Mesas-Lafarga P., Prats-Climent J., Rodríguez-Álvarez M.J.

There has been a strong interest in using neural networks to solve several tasks in PET medical imaging. One of the main problems faced when using neural networks is the quality, quantity, and availability of data to train the algorithms. In order to address this issue, we have developed a pipeline that enables the generation of voxelized synthetic PET phantoms, simulates the acquisition of a PET scan, and reconstructs the image from the simulated data. In order to achieve these results, several pieces of software are used in the different steps of the pipeline. This pipeline solves the problem of generating diverse PET datasets and images of high quality for different types of phantoms and configurations. The data obtained from this pipeline can be used to train convolutional neural networks for PET reconstruction.

García-Ortíz J.V., Mora M.C., Cerdá-Boluda J.

In the field of biomechanical modeling, the development of a prosthetic hand with dexterity comparable to the human hand is a multidisciplinary challenge involving complex mechatronic systems, intuitive control schemes, and effective body interfaces. Most current commercial prostheses offer limited functionality, typically only one or two degrees of freedom (DoF), resulting in reduced user adoption due to discomfort and lack of functionality. This research aims to design a computationally efficient low-level control algorithm for prosthetic hand fingers to be able to (a) accurately manage finger positions, (b) anticipate future information, and (c) minimize power consumption. The methodology employed is known as model-based predictive control (MBPC) and starts with the application of linear identification techniques to model the system dynamics. Then, the identified model is used to implement a generalized predictive control (GPC) algorithm, which optimizes the control effort and system performance. A test bench is used for experimental validation, and the results demonstrate that the proposed control scheme significantly improves the prosthesis’ dexterity and energy efficiency, enhancing its potential for daily use by people with hand loss.

Castillo D., Alejandro R.J., García S., Rodríguez-Álvarez M.J., Lakshminarayanan V.

Nowadays, the diagnosis of numerous diseases is facilitated by medical imaging. In that context, the identification of brain lesions presented as White Matter Hyperintensities (WHMs) and their related diseases is essential to have a correct diagnosis. Machine- and deep learning (subfields within artificial intelligence) could support the diagnosis (especially in complex medical images) by leveraging the structure and regularities within the imaging data. This project presents a technique for the classification of WHMs concerning ischemia and demyelination through the analysis of the region of interest (ROI) features of magnetic resonance images. To do that, we analyzed radiomic features using a combination of principal component analysis (PCA) and support vector machine (SVM) classification. Next, we used a transfer learning fine-tuned ResNet18 model to more thoroughly analyze and classify lesioned ROIs. For that, we used patient data alone and additional synthetic data (generated using spectral generative adversarial networks -SNGAN). The results show an accuracy mean value of 0.96 without data augmentation; while we had a value of 0.54 using synthetic data, a similar value was acquired with radiomics-informed SVM classification (0.56). These findings constitute a starting point for future projects exploring different ways of informing and fine-tuning artificial intelligence models to detect, classify, and segment MRI pathologies characterized by small lesions.

Vivanco Gualán R.I., del Cisne Jiménez Gaona Y., Castillo Malla D.P., Rodríguez-Alvarez M.J., Lakshminarayanan V.

Breast thermography captures infrared radiation images to monitor skin surface temperature changes non-invasively. This data, when combined with artificial intelligence, facilitates early breast cancer diagnosis and detection. However, training deep learning algorithms such as convolutional neural networks is challenging due to the limited number of images. The primary objective of this study is to create a set of synthetic breast thermographic images using segmentation and data augmentation techniques. In this work, we propose 1) Using public breast thermography databases, 2) Segmenting the region of interest with the U-Net network, 3) Increasing the variety of thermographic images using the SNGAN model, and 4) Evaluating the performance and accuracy of the previous algorithms with statistical metrics. The results indicate that the U-Net achieved an IoU of 0.96 and a Dice coefficient of 0.97. The SNGAN network generated 2000 synthetic images, reflected in a KID value of 4.54. In conclusion, U-Net is highly effective for segmenting regions of interest in thermographic images, and SNGAN shows promising results in synthetic image generation.

Rivero-Buceta E., Bernal-Gómez A., Vidaurre-Agut C., Lopez Moncholi E., María Benlloch J., Moreno Manzano V., David Vera Donoso C., Botella P.

Docetaxel (DTX) is a recommended treatment in patients with metastasic prostate cancer (PCa), despite its therapeutic efficacy is limited by strong systemic toxicity. However, in localized PCa, intratumoral (IT) administration of DTX could be an alternative to consider that may help to overcome the disadvantages of conventional intravenous (IV) therapy. In this context, we here present the first in vivo preclinical study of PCa therapy with nanomedicines of mesoporous silica nanoparticles (MSN) and DTX by IT injection over a xenograft mouse model bearing human prostate adenocarcinoma tumors. The efficacy and tolerability, the biodistribution and the histopathology after therapy have been investigated for the DTX nanomedicine and the free drug, and compared with the IV administration of DTX. The obtained results demonstrate that IT injection of DTX and DTX nanomedicines allows precise and selective therapy of non-metastatic PCa and minimize systemic diffusion of the drug, showing superior activity than IV route. This allows reducing the therapeutic dose by one order and widens substantially the therapeutic window for this drug. Furthermore, the use of DTX nanomedicines as IT injection promotes strong antitumor efficacy and drug accumulation at the tumor site, improving the results obtained with the free drug by the same route.

Contreras T., Palmeiro B., Almazán H., Para A., Martínez-Lema G., Guenette R., Adams C., Álvarez V., Aparicio B., Aranburu A.I., Arazi L., Arnquist I.J., Auria-Luna F., Ayet S., Azevedo C.D., et. al.

Abstract The NEXT-White detector, a high-pressure gaseous xenon time projection chamber, demonstrated the excellence of this technology for future neutrinoless double beta decay searches using photomultiplier tubes (PMTs) to measure energy and silicon photomultipliers (SiPMs) to extract topology information. This analysis uses 83mKr data from the NEXT-White detector to measure and understand the energy resolution that can be obtained with the SiPMs, rather than with PMTs. The energy resolution obtained of (10.9 ± 0.6)%, full-width half-maximum, is slightly larger than predicted based on the photon statistics resulting from very low light detection coverage of the SiPM plane in the NEXT-White detector. The difference in the predicted and measured resolution is attributed to poor corrections, which are expected to be improved with larger statistics. Furthermore, the noise of the SiPMs is shown to not be a dominant factor in the energy resolution and may be negligible when noise subtraction is applied appropriately, for high-energy events or larger SiPM coverage detectors. These results, which are extrapolated to estimate the response of large coverage SiPM planes, are promising for the development of future, SiPM-only, readout planes that can offer imaging and achieve similar energy resolution to that previously demonstrated with PMTs.

Toussaint M., Loignon-Houle F., Auger É., Lapointe G., Dussault J., Lecomte R.

Abstract Objective. Acollinearity of annihilation photons (APA) introduces spatial blur in positron emission tomography (PET) imaging. This phenomenon increases proportionally with the scanner diameter and it has been shown to follow a Gaussian distribution. This last statement can be interpreted in two ways: the magnitude of the acollinearity angle, or the angular deviation of annihilation photons from perfect collinearity. As the former constitutes the partial integral of the latter, a misinterpretation could have significant consequences on the resulting spatial blurring. Previous research investigating the impact of APA in PET imaging has assumed the Gaussian nature of its angular deviation, which is consistent with experimental results. However, a comprehensive analysis of several simulation software packages for PET data acquisition revealed that the magnitude of APA was implemented as a Gaussian distribution. Approach. We quantified the impact of this misinterpretation of APA by comparing simulations obtained with GATE, which is one of these simulation programs, to an in-house modification of GATE that models APA deviation as following a Gaussian distribution. Main results. We show that the APA misinterpretation not only alters the spatial blurring profile in image space, but also considerably underestimates the impact of APA on spatial resolution. For an ideal PET scanner with a diameter of 81 cm, the APA point source response simulated under the first interpretation has a cusp shape with 0.4 mm FWHM. This is significantly different from the expected Gaussian point source response of 2.1 mm FWHM reproduced under the second interpretation. Significance. Although this misinterpretation has been found in several PET simulation tools, it has had a limited impact on the simulated spatial resolution of current PET scanners due to its small magnitude relative to the other factors. However, the inaccuracy it introduces in estimating the overall spatial resolution of PET scanners will increase as the performance of newer devices improves.

Lamothe N., Andrés D., Carrión A., Camarena F., Pineda-Pardo J.A., Jiménez N.

Acoustic holograms can generate cavitation patterns of complex spatial distribution by shaping and steering the focal spot of therapeutic ultrasound systems. However, when monitoring these systems by passive cavitation detection, off-axis therapeutical targets and the receiver directivity may not be aligned. In this paper, we present passive cavitation beamforming to monitor a therapeutical ultrasound system using holograms targeted to arbitrary locations, in which both therapeutic and passive cavitation monitoring systems use 3D-printed acoustic lenses. The therapeutic system uses an acoustic hologram to focus the ultrasound beam on the target, which is off-axis. Then, a second lens is designed to beamform the cavitation signals which emerge from the therapeutic target, steering the directivity of the passive cavitation detector in the direction of the therapeutic focus and, in addition, compensating for skull aberrations. The system is experimentally tested with an ex-vivo macaque skull and a blood vessel phantom with microbubbles. In addition, results are compared with a standard confocal configuration and an off-axis configuration in the absence of the monitor lens. A parametric study is performed by varying the amplitude of the emitted signal and the impact on the behaviour of the microbubbles is analysed based on the cavitation index values. Results show that monitoring holograms align the passive cavitation detector response with the focal spot of the targeted therapeutic transducer. These holograms encode a fixed beamformer for cavitation signals in reception, increasing the sensitivity of cavitation emission at the target. In this way, cavitation doses can be used to locally monitor the cavitation activity of microbubbles, thus opening a new path to low-cost monitoring of therapeutic ultrasound systems.

Ocaña-Tienda B., Eroles-Simó A., Pérez-Beteta J., Arana E., Pérez-García V.M.

Abstract Background Lung nodules observed in cancer screening are believed to grow exponentially, and their associated volume doubling time (VDT) has been proposed for nodule classification. This retrospective study aimed to elucidate the growth dynamics of lung nodules and determine the best classification as either benign or malignant. Methods Data were analyzed from 180 participants (73.7% male) enrolled in the I-ELCAP screening program (140 primary lung cancer and 40 benign) with three or more annual CT examinations before resection. Attenuation, volume, mass and growth patterns (decelerated, linear, subexponential, exponential and accelerated) were assessed and compared as classification methods. Results Most lung cancers (83/140) and few benign nodules (11/40) exhibited an accelerated, faster than exponential, growth pattern. Half (50%) of the benign nodules versus 26.4% of the malignant ones displayed decelerated growth. Differences in growth patterns allowed nodule malignancy to be classified, the most effective individual variable being the increase in volume between two-year-interval scans (ROC-AUC = 0.871). The same metric on the first two follow-ups yielded an AUC value of 0.769. Further classification into solid, part-solid or non-solid, improved results (ROC-AUC of 0.813 in the first year and 0.897 in the second year). Conclusions In our dataset, most lung cancers exhibited accelerated growth in contrast to their benign counterparts. A measure of volumetric growth allowed discrimination between benign and malignant nodules. Its classification power increased when adding information on nodule compactness. The combination of these two meaningful and easily obtained variables could be used to assess malignancy of lung cancer nodules.