Recently, object detection in bioimage analysis has achieved interesting progress through deep learning models like those in the vision transformers, enabling high-accuracy predictions for tasks such as lesion detection in radiology, ophthalmology and organ segmentation. However, applying these models in healthcare raises critical challenges related to privacy and explainability. Centralized training is often infeasible due to strict data-sharing constraints, such as patient privacy regulations, while the black-box nature of modern detectors hinders clinician trust and adoption, particularly in clinical decision-making. To address these issues, we introduce BIOFED, a framework for privacy-preserving and explainable object detection specifically tailored for biomedical imaging. The idea behind BIOFED is to integrate distributed training of object detectors (for instance, the models belonging to the YOLO family) across decentralized clinical or research clients using aggregation strategies like FedAvg and FedProx, ensuring sensitive patient data never leaves the client side. Furthermore, BIOFED incorporates an explainability module based on Class Activation Mapping techniques, by generating heatmaps that highlight bioimage regions, like suspicious lesions in a chest X-ray or retinal fundus image that are most influential for model predictions. This combination of federated object detection and explainability enhances transparency, accountability, and trustworthiness in safety-critical bioimaging scenarios.

The lack of broad, high-quality, labeled datasets and, crucially, stringent data protection laws frequently make it difficult to integrate Deep Learning (DL), with medical imaging. These limitations make it impossible to centralize data, which results in models that don't generalize to other institutions (non-IID data). To address this, we introduce FLIP_VIT, a novel framework leveraging Federated Learning (FL) for explainable and privacy-preserving biomedical imaging. FL is a decentralized system where the training algorithm moves to the data. Hospitals (clients) train a model locally on their private data, and only the model updates (weights or gradients) are securely shared and aggregated by a central server using techniques like Federated Avareging (FedAvg) and Federeted Proximal (FedProx) to create a global model. Patient data never leaves its source, ensuring privacy and compliance.

Breast cancer is a kind of cancer that forms in the cells of the breasts. After skin cancer, breast cancer represents the most common cancer diagnosed in women in the United States. As a matter of fact, in January 2022, there are more than 3.8 million women with a history of breast cancer in the United States, this is the reason why there is a need for novel methods for automatic breast cancer screening, with the aim of starting any therapy as quickly as possible to try to limit the proliferation of the disease. In this article, we propose a method aimed at detecting breast cancer through a deep learning network developed by authors. Moreover, the proposed method is able to provide prediction explainability by means of class activation mapping, aimed to automatically highlight the suspicious area on the image. We take into account a way to understand whether the cancer prediction and localization can be considered robust by analyzing the output of two different class activation mapping algorithms. We evaluate the effectiveness of the proposed method by using a dataset composed of 9,016 images obtaining an accuracy equal to 93.5%, thus showing the effectiveness of the proposed network for breast cancer detection and localization.

Skin lesions are any abnormal growths or appearances on the skin, ranging from benign (i.e., non-cancerous) to malignant (i.e., cancerous). The identification of a skin lesion is a crucial task that is carried out in short periods of time to initiate an eventual therapeutic treatment. In this paper, we propose a method for automatic skin lesion detection, implementing Convolutional Neural Networks. Moreover, with the aim of providing a rationale behind the model prediction, we also consider explainability by adopting two different Class Activation Mapping algorithms, which highlight regions in skin images that most contribute to the network’s classification decision. We also include the indices of similarity for further quantitative analysis. Several Convolutional Neural Networks are considered, by obtaining the best results with the MobileNet model, achieving an accuracy equal to 0.935 in skin lesion detection. Moreover, in the experimental analysis, we discuss the effectiveness of Class Activation Mapping algorithms exploited for skin lesion localization.

Background and Objective: Alzheimer’s disease is nowadays the most common cause of dementia. It is a degenerative neurological pathology affecting the brain, progressively leading the patient to a state of total dependence, thus creating a very complex and difficult situation for the family that has to assist him/her. Early diagnosis is a primary objective and constitutes the hope of being able to intervene in the development phase of the disease. Methods: In this paper, a method to automatically detect the presence of Alzheimer’s disease, by exploiting deep learning, is proposed. Five different convolutional neural networks are considered: ALEX_NET, VGG16, FAB_CONVNET, STANDARD_CNN and FCNN. The first two networks are state-of-the-art models, while the last three are designed by authors. We classify brain images into one of the following classes: non-demented, very mild demented and mild demented. Moreover, we highlight on the image the areas symptomatic of Alzheimer presence, thus providing a visual explanation behind the model diagnosis. Results: The experimental analysis, conducted on more than 6000 magnetic resonance images, demonstrated the effectiveness of the proposed neural networks in the comparison with the state-of-the-art models in Alzheimer’s disease diagnosis and localization. The best results in terms of metrics are the best with STANDARD_CNN and FCNN with accuracy, precision and recall between 98% and 95%. Excellent results also from a qualitative point of view are obtained with the Grad-CAM for localization and visual explainability. Conclusions: The analysis of the heatmaps produced by the Grad-CAM algorithm shows that in almost all cases the heatmaps highlight regions such as ventricles and cerebral cortex. Future work will focus on the realization of a network capable of analyzing the three anatomical views simultaneously.

Data poisoning represents a set of techniques aimed at perturbing data for training machine learning models, affecting performance. Such intentional attacks are widespread in many applications involving deep learning algorithms and are aimed to provide misclassifications. In this paper, data poisoning on retinal images for the diabetic retinopathy binary classification (health and sick) is presented and evaluated. The presented attacks are almost imperceptible perturbations of the images that nevertheless decrement the metrics of the trained models. Once exposed to the data poisoning distortions on these images, a possible countermeasure to enhance the security from these attacks is shown. In this way, the robustness and vulnerabilities of the network are highlighted and the best result is also analyzed through the use of heatmaps, for the qualitative point of view. The paper aims to focus on the effects of data poisoning in deep learning model testing phase and to discuss possible countermeasures.

Cardiovascular diseases include a very long series of diseases that afflict many people in the world. Many of them can be diagnosed by listening to the heartbeat, however in the face of the large number of patients performing checks, great delays can occur given the few doctors available. In this paper we propose a convolutional neural network aimed to discriminate regular heartbeats from abnormal ones, making a first screening of patients. Moreover we provide classification explainability through activation maps. The experimental analysis consists of 3240 (regular and abnormal) heartbeat phonocardiogram signals, showing the effectiveness of the proposed method.

Often when we have a lot of data available we can not give them an interpretability and an explainability such as to be able to extract answers, and even more so diagnosis in the medical field. The aim of this contribution is to introduce a way to provide explainability to data and features that could escape even medical doctors, and that with the use of Machine Learning models can be categorized and "explained".

The new coronavirus disease (COVID-19), declared a pandemic on 11 March 2020 by the World Health Organization, has caused over 6 million victims worldwide. Because of the rapid spread of the virus, with the aim to perform screening we exploit deep learning model to quickly diagnose altered respiratory conditions. In this paper, we propose a method to recognize and classify cough audio files into three classes to distinguish patients with COVID-19 disease, symptomatic ones and healthy subjects, with the use of a convolutional neural network (CNN). Cough audios were recorded by using a smartphone and its built-in microphone. From cough recordings, we generate spectrogram images and we obtain an accuracy equal to 0.82 with a deep learning network developed by authors. Our method also provides heatmaps, which show the relevant input areas used by the model for the final forecast, and this aspect ensures the explainability of the method.

Diabetic retinopathy is a disease afflicting the retina and currently is manually diagnosed by specialists through eye tomography inspection. In order to assist the clinician in this time-consuming task, in this paper, we propose a method aimed to automatically diagnose the (proliferative and non-proliferative) diabetic retinopathy by exploiting deep learning. Furthermore, we investigate the possibility to automatically localise the areas related to the disease by exploiting class activation maps. We evaluate different deep learning models from a quantitative point of view (i.e, using metrics like accuracy, precision and recall) and a qualitative point of view (by exploiting class activation maps and image similarity metrics) with the aim to understand the quality of predictions performed by a model in retinopathy diagnosis, reducing the amount of knowledge required to assess the model performance. From the experimental analysis is emerging that deep learning shows an interesting diagnostic potential in the retinopathy disease localisation and can effectively help the clinician in retinopathy diagnosis. Moreover, the adoption of the class activation maps and its comparison evaluation can help the developers to debug the training step of the model without medical expertise.

Image segmentation is becoming a component of medical image processing with a significant role to analyze gross anatomy, to locate an infirmity and to plan the surgical procedures. Segmentation of brain magnetic resonance Imaging (MRI) is of increasing importance for the accurate diagnosis. For this reason, precise and accurate segmentation of brain MRI is a challenging task. In this paper we propose an approach for brain segmentation from MRI. Our method relies on a modified version of the U-Net convolutional neural network. Experiments performed on high grade brain cancer MRI demonstrate the effectiveness of the proposed approach for high grade brain cancer segmentation.