More than 10 scientific papers authored by INESC TEC researchers were submitted to the IEEE Engineering in Medicine and Biology Society (EMBC) Annual International Conference – one of the most important events in the field, attracting thousands of participants each year.
Widely regarded as the largest and most influential conference in Biomedical Engineering worldwide, EMBC welcomed INESC TEC contributions – with the acceptance of 10 papers involving researchers from the institute, marking the solid Portuguese presence at this year’s edition, held in Copenhagen, Denmark.
Inês Lopes, author of the paper Image Classification: On the Impact of Input Resolution on CNN-based Gastrointestinal Endoscopy, highlighted the “extremely inspiring” experience of participating in a conference that showcased “the latest advances in Artificial Intelligence (AI) applied to medicine”, while also “encouraging international collaboration”.
Within medical imaging, the area most closely connected to her research, “there was a clear trend towards self-learning techniques for image segmentation”. “This approach reduces dependence on manually annotated data, which is an important step forward for the efficiency of diagnostic support systems,” she explained.
Her paper explores how the resolution of endoscopic images affects the performance of convolutional neural networks (CNNs) in detecting intestinal metaplasia – a precursor lesion of gastric cancer. To ensure robustness and multiple perspectives, two clinical datasets with different image resolutions and modalities were used.
According to Inês Lopes, “the study showed that resolutions higher than the conventional 224×224 significantly improved model performance in detecting gastric lesions, as they preserve essential visual detail”. “These results reinforce the importance of original image quality in training more effective clinical decision-support systems.”
Another example of science applied to medicine is the paper Language Model of Lung Nodules in LNDb Medical Reports, part of a broader line of work that has already motivated a patent proposal. The technology “allows linking descriptions of lung nodules in medical reports to automatically locate said nodules in CT scans”, explainws Miguel Coimbra. This work focuses specifically on the textual analysis algorithms used.
Miguel Coimbra is also a co-author of Predicting Endoscopic Grading of Gastric Intestinal Metaplasia using Small Patches – a collaboration between INESC TEC and the Instituto Português de Oncologia do Porto. “We demonstrate the first viable deep-learning-based solution for characterising EGGIM – an endoscopic grading system – using data collected through this partnership,” he stated.
This is the first time an EGGIM estimate has been validated, showing that the models can identify high-risk patients with sensitivity comparable to that of expert clinical staff. According to the researcher, “because this pathological phenomenon is visually dominated by texture, we can process only parts of the image, simplifying model training”.
In Smart Vest and Metrics for Physical Education using ECG & IMU, INESC TEC researchers Miguel Velhote and Rafael Aguiar describe work developed within the Texp@ct project. “We detail sensor-fusion and signal-processing algorithms for a smart garment, presenting preliminary results on assessing fundamental movements and exercise intensity for physical education and youth sports,” explained Miguel Velhote. “In school sports, objective analysis of physiological and biomechanical metrics offers a highly relevant tool for assessing and monitoring adolescents. Indicators like heart rate, activity state and exercise intensity, among others, are already in testing and validation stages within the Texp@ct solution,” added Rafael Aguiar.
Francisco Vieira and Joshua Woods focused on developing a reference architecture for synchronising multiple biomedical wearable devices using Bluetooth Low Energy (BLE). In WeSync(BLE): A Reference Synchronization Architecture of Multiple Wearable BLE-Based Biomedical Devices, the researchers highlight the importance of “reliable synchronisation across different wearables for accurate analysis of physiological signals”.
“The work will have a significant impact on our research lines, particularly in Parkinson’s disease and foetal monitoring,” they mentioned. The paper stems from Francisco Vieira’s PhD work.
Rui Castro, INESC TEC researcher, addressed coronary artery disease – the most common form of cardiovascular disease. Coronary artery calcium scoring using CT remains the most reliable diagnostic indicator. Although deep-learning models exist for coronary calcium segmentation, their interpretation is limited by the “black-box” nature of such models. In Contrastive Coronary Artery Calcification Image Retrieval in Computed Tomography, the authors propose a supervised contrastive retrieval pipeline that enhances interpretability by providing visually similar examples of coronary calcification.
Besides Biomedical Engineering, researchers in Telecommunications and Multimedia had five additional papers accepted.
In Robust Visual Transformers for Medical Image Classification, João Montrezol, Hugo S. Oliveira, Jorge Araújo and Hélder Oliveira explore the boundaries of current architectures, focusing on developing new techniques for highly complex tasks like medical imaging, which often involves high variability, class imbalance and limited sample sizes. “We propose a combined set of regularisation and data-augmentation techniques to improve model performance,” said Hélder Oliveira.
The new techniques include a modified loss function and a smoothly differentiable activation function, enabling “more stable training and improved model performance”. “The results show that incorporating these techniques enhances both performance and training convergence,” he added.
Limitations in current methods for predicting aesthetic outcomes in breast cancer treatment (as well as the constraints of existing datasets) motivated Conditional Generative Adversarial Network for Predicting the Aesthetic Outcomes of Breast Cancer Treatment, authored by Helena Montenegro, Maria J. Cardoso and Jaime S. Cardoso.
The researchers developed a mask-guided Generative Adversarial Network (GAN) capable of manipulating breast shape and nipple position in mammographic imagery based on masks indicating the desired outcome. “We used an auxiliary segmentation network to guide the GAN in producing the desired changes during training,” they explained. Experiments on a private dataset of post-treatment breast aesthetics confirmed the model’s ability to manipulate shape and outperformed state-of-the-art methods.
A lack of generalisation to out-of-domain data motivated Domain-Specific Data Augmentation for Lung Nodule Malignancy Classification by Margarida Gouveia, Jorge Araújo, Hélder P. Oliveira and Tânia Pereira. “We propose a ResNet architecture with 2.5D inputs capable of preserving spatial information through three anatomical planes,” the team explained. Domain-specific data augmentation designed for CT scans was also applied.
Chest CT scans are essential for diagnosing pulmonary anomalies, including lung cancer. However, their use in deep-learning training is “often limited by data scarcity, annotation cost and privacy concerns”.
In Conditional Score-based Diffusion Models for Lung CT Scans Generation, António F. Cardoso, Pedro Sousa, Hélder P. Oliveira and Tânia Pereira explore “score-based diffusion models for conditional generation of lung CT slices”. Two scenarios were tested: “one using only lung segmentation masks, and another incorporating both lung- and nodule-segmentation maps”.
Finally, Phenotypic Characterization of Sleep Apnea Using Clusters Derived from Subject-Based SpO₂ Weighted Correlation Networks introduces a new method using SpO₂-weighted correlation networks and modularity analysis to identify clinically relevant subgroups of patients with Obstructive Sleep Apnoea (OSA). The study used data from 2,641 participants in the Sleep Heart Health Study, from which 43 SpO₂-related variables were extracted from polysomnography recordings.
“The robustness of the results was ensured through a bootstrap procedure, and subgroup identification was performed using the Blondel modularity algorithm without requiring a predefined number of groups,” concluded researcher Daniela Ferreira Santos.
The researcher mentioned in this news piece are associated with INESC TEC, the Faculty of Engineering of the University of Porto, the Faculty of Sciences of the University of Porto and the Faculty of Medicine of the University of Porto.
News, current topics, curiosities and so much more about INESC TEC and its community!
