IbPRIA 2025: 12th Iberian Conference on Pattern Recognition and Image Analysis

Coimbra, Portugal. June 30 - July 3, 2025

IbPRIA 2025 Accepted Papers

Oral Session 6 - Biomedical applications 2

Continual Deep Active Learning for Medical Imaging: Replay-Base Architecture for Context Adaptation

Rui Daniel, M. Rita Verdelho, Catarina Barata, Carlos Santiago

Abstract:
Deep Learning for medical imaging faces challenges in adapting and generalizing to new contexts. Additionally, it often lacks sufficient labeled data for specific tasks requiring significant annotation effort. This work proposes a novel Continual Active Learning (CAL) framework, called Replay-Base Architecture for Context Adaptation (RBACA), for robust medical image analysis. The Continual Learning (CL) part tackles adaptability and generalizability by enabling lifelong learning from a data stream while mitigating forgetting of previously learned knowledge. On the other hand, the Active Learning (AL) part reduces the number of required annotations for effective training. Based on the automatic recognition of shifts in image characteristics, RBACA employs a CL rehearsal method to continually learn from diverse contexts, and an AL component to select the most informative instances for annotation. We show that RBACA works in domain and class-incremental learning scenarios, by assessing its performance on the segmentation and diagnosis of cardiac images. The results show that RBACA outperforms a baseline framework without CAL, and a state-of-the-art CAL method across various memory sizes and annotation budgets. Our code is available in https://github.com/RuiDaniel/RBACA.

Informed decision making strategy for resampling in pain assessment

Miguel Carvalho, Daniela Pais, Raquel Sebastião, Armando Pinho, Susana Brás

Abstract:
Class imbalance can significantly impact the performance of learning algorithms, often leading to prediction bias toward the majority class. This challenge is particularly critical in healthcare-related domains, as medical datasets are often imbalanced, hindering the accurate prediction of the minority class, which is commonly the class of interest. As such, this work introduces a novel resampling algorithm, designated Genetic Beta Oversampling, which integrates user-defined preferences into the synthetic data generation process, allowing fine control over the model’s inclination towards false negatives or false positives. These user preferences are encoded in the form of a parameter, ß, which dictates the trade-off between recall and precision that the method should seek to achieve. This flexibility is particularly relevant in clinical settings, where prioritizing recall can enhance patient care by reducing missed diagnoses. We validate the approach using the EMPA dataset for the classification of pain induced by a cold stimulus, where prioritizing recall is essential to minimize missed pain detections. Experimental results demonstrate that our method outperforms SMOTE, SMOTE-IPF, and four cost-sensitive classifiers in terms of Fß-score across diverse ß settings. These findings underscore the method’s adaptability in recall-sensitive applications, such as pain assessment and clinical decision-making.

MIL vs. Aggregation: Evaluating Patient-Level Survival Prediction Strategies Using Graph-Based Learning

Maria Rita Verdelho, Alexandre Bernadino, Catarina Barata

Abstract:
Oncologists often rely on a multitude of data, including whole-slide images (WSIs), to guide therapeutic decisions, aiming for the best patient outcome. However, predicting the prognosis of cancer patients can be a challenging task due to tumor heterogeneity and intra-patient variability. While AI models have started to gain relevance in this field, they are challenged by the amount of available data, including the existence of multiple WSIs, each capturing different tumor regions and that may not be equally relevant. This raises a fundamental question: Should we use all WSIs to characterize the patient, or should we identify the most representative slide for prognosis? Our work seeks to answer this question by performing a comparison of various strategies for predicting survival at the WSI and patient level. The former treats each WSI as an independent sample, mimicking the strategy adopted in other works, while the latter comprises methods to either aggregate the predictions of the several WSIs or automatically identify the most relevant slide using multiple-instance learning (MIL). Additionally, we evaluate different Graph Neural Networks architectures under these strategies. We conduct our experiments using the MMIST-ccRCC dataset, which comprises patients with clear cell renal cell carcinoma (ccRCC). Our results show that MIL-based selection improves recall, suggesting that choosing the most representative slide benefits survival prediction.

Prediction of 30-day hospital readmission with clinical notes and EHR information

Tiago Almeida, Plinio Moreno, Catarina Barata

Abstract:
High hospital readmission rates are associated with significant costs and health risks for patients. Therefore, it is critical to develop predictive models that can support clinicians to determine wether or not a patient will return to the hospital in a relatively short period of time (e.g, 30-days). Nowadays, it is possible to collect both structured (electronic health records - EHR) and unstructured information (clinical notes) about a patient hospital event, all potentially containing relevant information for a predictive model. However, their integration is challenging. In this work we explore the combination of clinical notes and EHRs to predict 30-day hospital readmissions. We address the representation of the various types of information available in the EHR data, as well as exploring LLMs to characterize the clinical notes. We collect both information sources as the nodes of a graph neural network (GNN). Our model achieves an AUROC of 0.72 and a balanced accuracy of 66.7%, highlighting the importance of combining the multimodal information.

Publisher

Endorsed by

Technical Sponsors