This paper presents a set of technologies developed to exploit Grid infrastructures for breast cancer CAD, that include (1) federated repositories of mammography images and clinical data over Grid storage, (2) a workstation for mammography image analysis and diagnosis and (3) a framework for data analysis and training machine learning classifiers over Grid computing power specially tuned for medical image based data. An experimental mammography digital repository of approximately 300 mammograms from the MIAS database was created and classifiers were built achieving a 0.85 average area under the ROC curve in a dataset of 100 selected mammograms with representative pathological lesions and normal cases. Similar results were achieved with classifiers built for the UCI Breast Cancer Wisconsin dataset (699 features vectors). Now these technologies are being validated in a real medical environment at the Faculty of Medicine in Porto University after a process of integrating the tools within the clinicians workflows and IT systems.

Breast cancer is one of the most diagnosed types of cancer. Early detection of breast cancer is the best practice to reduce its high mortality rate. Double reading is a common practice to avoid missed cancer cases during the lecture of mammograms. Due to the high cost of this technique and the lack of certified radiologists, computer-aided diagnosis systems have become very popular in the last decade. However, this technique is not perfect, and has some significant disadvantages, like higher recall rates.

Artificial intelligence (AI) systems are increasingly used in medical imaging for disease diagnosis, yet their vulnerability to adversarial attacks poses significant risks for clinical deployment. In this work, we systematically evaluate the susceptibility of VolBrain, a widely used third-party neuroimaging diagnostic platform, to universal black-box adversarial attacks. We generate adversarial perturbations using a surrogate convolutional neural network trained on a different dataset and with a different architecture, representing a worst-case scenario for the attacker where they have no access to the internals of the system. For this, we employ both the Fast Gradient Sign Method (FGSM) and DeepFool attacks. Our results show that these perturbations can reliably degrade the diagnostic performance of VolBrain, with DeepFool-based attacks being particularly effective for comparable perturbation sizes. We further demonstrate that a simple Mean Attack approach is also effective in degrading VolBrain performance, showing that this system is vulnerable to universal attacks, that is, perturbations agnostic to the input. These findings highlight the substantial risk posed by universal black-box adversarial attacks, even when attackers lack access to the target model or its training data. Our study underscores the urgent need for robust defense mechanisms and motivates further research into the adversarial robustness of medical AI systems.

Multi-modal models that fuse neuroimaging with clinical assessment data represent the current state of the art for automated Alzheimer's disease detection, yet their adversarial robustness remains poorly understood. We systematically investigated adversarial vulnerability in CogniNetMM, a model that fuses 3D structural MRI volumes with neuropsychological clinical variables, using the Fast Gradient Sign Method and DeepFool in three modality configurations: MRI only, clinical variables only, and both jointly. We further introduced a mean attack framework, in which the average perturbation vector across samples serves as a fixed-direction probe of the decision boundary, decoupling the contribution of attack direction from that of sample position.During training, we identified modality collapse, a training instability in which the fusion layer progressively suppresses the MRI pathway. Collapse probability decreased with larger batch sizes and was further reduced by stratified sampling on class-imbalanced data. Across all clinical variable configurations and both attack methods, the joint multi-modal attack achieved higher success than the average of the two unimodal attacks. For DeepFool, the advantage was strong enough that the joint attack outperformed each unimodal attack individually near the decision boundary. The mean attack replicated this result under a fixed perturbation direction, confirming that the advantage is a structural property of the fused decision boundary rather than an artifact of per-sample gradient alignment. These findings are consistent with a concave original-class region in the joint input space: non-linear modality coupling creates adversarially reachable regions that neither modality perturbation can access independently. Together, these results show that heterogeneous data fusion introduces emergent adversarial vulnerabilities beyond what unimodal analysis predicts, and that standard training practices on imbalanced medical datasets carry a risk of silent modality suppression.