The number of patients with inflammatory bowel disease (IBD) has been increasing. The diagnosis is a difficult task for the gastroenterologist performing the endoscopic examination. However, in order to prescribe medical treatment and provide quality of life to the patient, the diagnosis must be quick and accurate. This paper presents a study where the objective is to collect and analyse endoscopic images referring to Crohn’s disease and Ulcerative colitis using four deep neural networks: ResNet50, InceptionV3, VGG16, and a hybrid model. The hybrid model consists of the combination of two architectures, a CNN and an LSTM. The main focus is on the understanding of the networks, offering through this paper a comparative study of five interpretability models, Grad-CAM, LIME, SHAP values, RISE, and Occlusion sensitivity. The obtained results demonstrate that it is possible to automate the process of diagnosing patients with IBD using deep networks for processing images collected during an endoscopic examination. Thus, we can develop tools that, with the aid of interpretability models, assist medical specialists in diagnosing the disease by understanding the specific region of the mucosa the network considered when making a decision.

Over the past decade, there has been growing evidence that artificial intelligence and radiomics may be helpful in the prediction of clinical outcomes in the entire prostate cancer disease continuum, such as the prediction of disease aggressiveness. However, the radiomics pipeline's dependence on segmentation masks has made it challenging to build machine-learning algorithms robust to inter- and intra-radiologist segmentation variability.

With the goal of getting insight into the best methodology to build models that are robust to this heterogeneity, two radiologists were asked to draw whole prostate gland segmentations on T2W and DWI MRI examinations, and the resulting radiomic features calculated were used in several model training approaches: training with purely stable radiomic features according to their intraclass correlation coefficient (ICC); training independently with features extracted from each radiologist's mask; training with the feature average between both radiologists; extracting radiomic features from the intersection or union of the two masks; and creating a heterogeneous dataset by randomly selecting one of the radiologists' masks for each patient.

The classifier trained with this last resampled dataset presented with the lowest generalization error, suggesting that training with heterogeneous data leads to the development of the most robust classifiers. On the contrary, removing features with low ICC resulted in the highest generalization error.

We present a novel hybrid framework that allows for joint learning of a Hidden Markov Chain and Artificial Neural Network in the context of fundamental heart sound segmentation. The Markovian nature of the model allows for unsupervised end-to-end training and our experiments reveal improvement of up to 3.90\% Positive Predictive Value over a pre-trained baseline using the PhysioNet 2016 and 2022 datasets.

Gastric Cancer (GC) and Colorectal Cancer (CRC) are some of the most common cancers in the world. The most common diagnostic methods are upper endoscopy and biopsy. Possible expert distractions can lead to late diagnosis. GC is a less studied malignancy than CRC, leading to scarce public data that difficult the use of AI detection methods, unlike CRC where public data are available. Considering that CRC endoscopic images present some similarities with GC, a CRC Transfer Learning approach could be used to improve AI GC detectors. This paper evaluates a novel Transfer Learning approach for real-time GC detection, using a YOLOv4 model pre-trained on CRC detection. The results achieved are promising since GC detection improved relatively to the traditional Transfer Learning strategy.

Cardiovascular diseases (CVDs) are the primary global cause of death, imposing substantial clinical, and economic burdens. Accurate risk stratification tools are crucial for guiding clinical decisions and preventive care. In this study, we’ve employed Machine Learning techniques to integrate inflammation biomarkers with well-established Acute Coronary Syndrome (ACS) risk factors, to enhance the GRACE stratification tool. The developed approach combines clinical knowledge with data-driven techniques, ensuring interpretability and personalization without compromising performance. To validate our approach, we collaborated with the Cardiology Unit of Coimbra Hospital and University Centre (CHUC) and analyzed a dataset of 1544 ACS patients. Our approach improved ACS Risk Scores by 5% when compared to the widely used GRACE, offering clinicians a comprehensive and personalized tool to make informed decisions and provide better patient care.

Breast tumor analysis is essential to diagnose breast cancer. Accurate 3D segmentation of breast tumors from medical images is essential for comprehensive disease analysis. This paper presents an automated pipeline for segmenting multiple breast tumors from Magnetic Resonance Imaging (MRI) scans. Utilizing a 3D region growing algorithm, the study addresses challenges in detecting multiple tumors and seed point selection.

Successful segmentation of four tumors highlights the potential of this approach for automatic multi-tumor segmentation, suggesting compatibility with a classification model based on the segmented structural features.

Motion estimation in echocardiography is critical when assessing heart function and calculating myocardial deformation indices. Nevertheless, there are limitations in clinical practice, particularly with regard to the accuracy and reliability of measurements retrieved from images. In this study, deep learning-based motion estimation architectures were used to determine the left ventricular longitudinal strain in echocardiography. Three motion estimation approaches, PWC-Net, RAFT and FlowFormer, were applied to a simulated echocardiographic dataset, achieving an average end point error of 0.24, 0.22 and 0.21 mm per frame, respectively. Thus,

optical flow-based motion estimation has the potential to facilitate the use of strain imaging in clinical practice.

Point-of-care lung ultrasonography (LUS) plays a vital role in rapid lung assessment. Despite its utility in the clinical evaluation of pulmonary diseases, its use is currently limited due to the lack of qualified professionals to interpret these images. Therefore, this study harnesses content-based image retrieval (CBIR) techniques for LUS examinations, focusing on binary and multi-label classifications. Two pre-trained models were used for binary and multi-label classification, and feature-based retrieval was carried out. The used models demonstrate 88.3% binary retrieval accuracy, while achieving 65.6% accuracy for multi-label retrieval. Promising results have been achieved, but there remains a wealth of opportunities to explore, and it is of utmost importance to do so, in order to be able to accurately assist a health professional in their decision-making.

Using medical images from the rectus femoris muscle acquired through ultrasound and elastography, the present work analyses the classification of sarcopenia using modern deep learning architectures and conventional machine learning models. The dataset, consists of 180 medical images collected from 30 people with ages ranging from 20 to 75.

The study explores a variety of models, including deep learning models like DenseNet 121, VGG16, VGG19, ResNet50, and Inception V3, as well as traditional models like logistic regression and neural networks.

The performance of the neural network model is in line with deep learning models. The Neural Network achieved the best performance with an F1_score of 99.81%. This work provides insights into the function of classical and deep learning methodologies as diagnostic tools for early intervention and enhanced care of an aging global population, shedding light on their potential to distinguish sarcopenia properly.

Thyroid nodules, despite appearing as a discrete lesion, constitute a prevailing pathological occurrence within the global population. The timely detection and diagnosis can help preventing the pathology from growing, minimizing more severe effects on the human body. In this study, supervised machine learning and deep learning techniques were implemented to analyse ultrasound and elastography medical images increasing and improving the effectiveness of thyroid nodule detection. The results achieved using deep learning were superior to those achieved using machine learning. Specifically, for machine learning it was obtained a F1-Score of 97.20%, for the ultrasound images and a F1-Score of 75.40% using elastography images. Deep learning methodologies reached a F1-Score of 98.85% for ultrasound images and 89.15% for elastography images.

Segmentation of skeletal muscles in Magnetic Resonance Images (MRI) is essential for the study of muscle physiology and diagnosis of muscular pathologies. However, manual segmentation of large MRI volumes is a time-consuming task. The state-of-the-art on algorithms for muscle segmentation of MRI is still not very extensive and, more recently, uses mostly learning-based solutions, which require large amounts of data. This work proposes an automated segmentation method for Dixon scans of the thigh based on pixel-wise classification of local texture features using AdaBoost. The descriptor includes features from the Histogram of Oriented Gradients (HOG), Haar Wavelet filtering, and statistical measures from both the original MRI and the Laplacian of Gaussian (LoG) filtering. An atlas-based approach is then applied to the resulting muscle tissue segmentation to provide individual muscle labeling.

Recent research indicates a connection between epicardial adipose tissue (EAT) and Coronary Artery Disease (CAD). EAT is a type of fat situated within the pericardium, a thin membrane sac that covers the heart. Hence, its segmentation and quantification could prove valuable for investigating its potential as a CAD risk stratification tool. However, manually segmenting these structures proves to be a demanding and time-consuming task, making it unsuitable for clinical settings. This has driven the development of automated segmentation methods. This study introduces an automated method for segmenting EAT in CT scans. A U-Net framework is thus used to segment the pericardium, which then allows to segment the EAT through thresholding. The quantification metrics resulted in a bias of 0.98 ± 15.351 cm³ and a Pearson Correlation Coefficient (PCC) of 0.924. In terms of segmentation metrics, the values for DSC, recall, and precision were 0.749 ± 0.051, 0.766 ± 0.069, and 0.748 ± 0.085, respectively. The results indicate that satisfactory performance can be attained on an external dataset encompassing diverse anatomical variations, using solely public datasets for training. However, incorporating more data will enhance the robustness of this approach, particularly in outlier cases. Future approaches should prioritize refining the integration of 3D information to achieve a more precise segmentation, mainly on the lower pericardium.

Blood vessel segmentation in 3D medical images is important in several clinical practices, and its automation has been studied during the last decades. However, the current algorithms are still prone to segmentation errors, such as missing segments or an inadequate merging or splitting of branches. These errors may severely change the topology of the network, risking the success of future medical interventions that depend on it. This paper presents a review of the state-of-the-art machine learning algorithms for the automatic segmentation of blood vessels in 3D medical images, while also giving special focus to the topology coherence of the segmented blood vessel networks.

Quality control of medical images plays an important role in digital pathology since verifying that the images meet all requirements can imply manual analysis. Manual assessment of pathology specimen fragments is intended to ensure that the number of fragments described in the macroscopic report corresponds to the number of fragments present on the slide, avoiding the loss of valuable material during grossing. However, this process is currently performed manually and is time-consuming and subjective. We applied an object detection model, YOLOv5, to detect fragments and sets in Whole Slide-Images dataset from IMP Diagnostics. Subsequently, we counted the final number of fragments by dividing the number of fragments by the number of sets. We decided to add a reject option when the confidence was low, based on the value of the division of fragments and sets, forcing the rejection of the sample if this number is not an integer. When tested on a set of 700 images, the model achieves an overall accuracy of 87.9% (without rejection), which increases to 92.8% if we reject 10.9% of the samples. This approach can overcome the limitations of the manual process, improving the workflow and final diagnosis.

COVID-19 disease caused a recent pandemic period, due to its ease of transmission and high number of cases of infections. The pandemic had severe consequences for the mortality and morbidity of populations, especially the elderly. Several research projects have raised from the pandemic, such as the Predictive Models of COVID-19 Outcomes for Higher Risk Patients Towards a Precision Medicine (PREMO) project. In this paper, we report the development of a Web application providing a dashboard with COVID-19 statistical data analysis from patients of the six waves of COVID-19, in Portugal. The visualizations provided by the platform allow specific analyzes of the clinical profile for patients hospitalized in Intensive Care Units (ICU), for each wave. The dashboard is an useful tool to extract information and scientific knowledge about the disease evolution, providing also beneficial insights for future pandemics.

Metaplasia detection in upper gastrointestinal endoscopy is crucial to identify patients at higher risk of gastric cancer. Deep learning algorithms can be useful for detecting and localising these lesions during an endoscopy exam. However, a lot of annotated data is needed to train these models, which can be a problem in the medical field. To overcome this, data augmentation techniques are commonly applied to increase the dataset's variability but must be adapted to the specificities of the application scenario. In this study, we discuss the potential benefits and identify four key research challenges of a promising data augmentation approach: image combination methodologies, such as CutMix, for metaplasia detection and localisation in gastric endoscopy imaging modalities.

Adult-type diffuse gliomas represent the predominant malignant neoplasms within the central nervous system. The advent of targeted therapeutic options has amplified the allure of molecular biomarkers, directly impacting on selection of appropriate interventions. Nevertheless, the manual assessment process within pathological laboratories is burdened by time intensiveness and error susceptibility. In order to surmount this constraint, multimodal fusion models have been explored. These models aim to identify the two pivotal molecular biomarkers (IDH1 mutation and 1p/19q codeletion) within gliomas by harnessing the synergy of MRI and digital pathology examinations.

Cancer results from the accumulation of mutations in the genetic material of the cells. Furthermore, due to its evasive nature and sometimes rapid development, it is one of the deadliest diseases worldwide. Among the different types of cancer, lung cancer is one of the deadliest, being responsible for millions of deads worldwide. However, this alarming numbers can be reduced if the lung cancer is detected in its early stages. Therefore, there is a great interest in the scientific community to develop new early detection tools for lung cancer. In this paper, we access two main approaches for the detection of lung cancer in CT images: One Stage Detectors and Two Stage Detectors. Our results show that One Stage Detectors are able to achieve better performance than the Two Stage Detectors.

The human microbiome has garnered significant interest due to emerging evidence of its association with various diseases, notably cancer. Technological breakthroughs in DNA sequencing have played a pivotal role in enabling extensive research on the microbiome. However, to fully comprehend the intricate relationship between microbiome composition and cancer, the use of sophisticated data-analytical tools has become imperative. This study aimed to develop a machine learning-based approach to distinguish cancer types based on tissue-specific microbial information, using Random Forest algorithms and samples from The Cancer Microbiome Atlas database. Promising performances were achieved for head and neck, stomach, and colon cancer classification, with colon cancer accuracy exceeding 90% across the studies. However, distinguishing esophageal and rectum cancers from the remaining proved challenging. The findings suggest that anatomically adjacent cancers are more complex to identify due to microbial similarities. Despite limitations, employing machine learning for microbiome data analysis could lead to innovative strategies for improving cancer detection, prevention, and reducing disease burden.

Deep Learning (DL) is the go-to solution in many Computer Vision (CV) tasks, including Haematoxylin and Eosin (H&E)-stained Whole Slide Image (WSI) analysis. However, DL models usually generalize poorly to out-of-distribution (o.o.d.) samples. In computational pathology (CPath), this is a known limitation, in spite of the several strategies that have been proposed to attenuate this challenge and improve generalization. In this work, we propose data augmentation to learn more robust representations of H&E-stained WSIs for colorectal dysplasia grading and demonstrate that data augmentations tailored to CPath could be useful to generalize DL models to novel environments.

When making lung cancer diagnosis, physicians usually take into account data from different modalities, and artificial-intelligence based methods could follow the same approach, in order to allow a more comprehensive analysis. Nonetheless, a great proportion of related works focus solely on imaging data. This work intended to investigate the potential of different data sources for lung cancer classification. A ResNet18 network was trained to classify 3D CT nodule regions of interest~(ROI), and a random forest algorithm was used to classify clinical data. Intermediate and late fusion methodologies were also developed, that combined the information from clinical data and 3D CT nodule ROIs. The best result, an AUC of 0.8021, was achieved by an intermediate fusion model -- a fully connected layer that receives deep imaging features, obtained from a ResNet18 inference model, and clinical data. Lung cancer is a complex disease, and this study shows that the combination of distinct modalities may have the potential to allow a comprehensive analysis of the pathology.

Deep learning methodologies have been very successful in a large number of tasks, sometimes surpassing human performance. One of the most simple neural network architectures is the multi-layer perceptron (MLP) which tries to mimic the brain in some ways. These methodologies are generally trained using gradient descent as they are differentiable. In recent years, a new methodology called spiking neural networks (SNNs) was proposed. These networks can be seen as a more biologically realistic approach than artificial neural networks (ANNs), as they use discrete spikes to transmit information, instead of continuous values. In this paper, we study these networks, and present results achieved by training these models. We also evaluate several information theory quantities such as entropy and mutual information of these networks to extract the relationship between the inputs and the outputs.

Breast cancer conservative treatment (BCCT) is a form of Breast Cancer treatment used to treat early-stage breast cancer patients as an alternative to mastectomy. This procedure consists of removing the cancerous tissue and maintaining the remaining healthy tissue intact, thus, improving aesthetic results. Currently, there is still no gold standard for evaluating the aesthetic outcome of BCCT objectively.

Recent approaches, such as the BCCT.core, automatically classify the results based on key features extracted from digital photographs of the breast. For this automatic approach to be applicable on a large scale it is necessary to automatically extract breast features from digital photographs. Our approach improves upon previous work that used the shortest path for the detection of breast contour by using a Deep Learning model for edge detection. This method achieved state-of-the-art results and surpassed the conventional method on 2 out of 3 datasets.

This paper presents a novel application of the YOLOv8 model for addressing the task of detecting organelles in Focused Ion Beam Scanning Electron Microscopy (FIB-SEM) biological images. The motivation behind this research stems from the need for accurate and efficient detection techniques to identify complex microstructures in biological material samples. The primary objective of this project involves fine-tuning of YOLOv8, a state-of-the-art object detection model, for semantic detection on FIB-SEM images. The study involves retraining the YOLOv8 model to suit the unique characteristics of FIB-SEM data, optimizing its performance for precise delineation of cellular structures. Experimental results showcase the model’s performance in effectively detecting intricate features while maintaining computational efficiency, enabling in-depth analysis and understanding of diverse microscopy samples in scientific research and industrial applications.