Background: Cutaneous melanoma is the most aggressive form of skin cancer, responsible for most skin cancer-related deaths. Recent advances in artificial intelligence, jointly with the availability of public dermoscopy image datasets, have allowed to assist dermatologists in melanoma identification. While image feature extraction holds potential for melanoma detection, it often leads to high-dimensional data. Furthermore, most image datasets present the class imbalance problem, where a few classes have numerous samples, whereas others are under-represented. Methods: In this paper, we propose to combine ensemble feature selection (FS) methods and data augmentation with the conditional tabular generative adversarial networks (CTGAN) to enhance melanoma identification in imbalanced datasets. We employed dermoscopy images from two public datasets, PH2 and Derm7pt, which contain melanoma and not-melanoma lesions. To capture intrinsic information from skin lesions, we conduct two feature extraction (FE) approaches, including handcrafted and embedding features. For the former, color, geometric and first-, second-, and higher-order texture features were extracted, whereas for the latter, embeddings were obtained using ResNet-based models. To alleviate the high-dimensionality in the FE, ensemble FS with filter methods were used and evaluated. For data augmentation, we conducted a progressive analysis of the imbalance ratio (IR), related to the amount of synthetic samples created, and evaluated the impact on the predictive results. To gain interpretability on predictive models, we used SHAP, bootstrap resampling statistical tests and UMAP visualizations. Results: The combination of ensemble FS, CTGAN, and linear models achieved the best predictive results, achieving AUCROC values of 87% (with support vector machine and IR=0.9) and 76% (with LASSO and IR=1.0) for the PH2 and Derm7pt, respectively. We also identified that melanoma lesions were mainly characterized by features related to color, while not-melanoma lesions were characterized by texture features. Conclusions: Our results demonstrate the effectiveness of ensemble FS and synthetic data in the development of models that accurately identify melanoma. This research advances skin lesion analysis, contributing to both melanoma detection and the interpretation of main features for its identification.

Machine learning (ML) models have been extensively used for tabular data classification but recent works have been developed to transform tabular data into images, aiming to leverage the predictive performance of convolutional neural networks (CNNs). However, most of these approaches fail to convert data with a low number of samples and mixed-type features. This study aims: to evaluate the performance of the tabular-to-image method named low mixed-image generator for tabular data (LM-IGTD); and to assess the effectiveness of transfer learning and fine-tuning for improving predictions on tabular data. We employed two public tabular datasets with patients diagnosed with cardiovascular diseases (CVDs): Framingham and Steno. First, both datasets were transformed into images using LM-IGTD. Then, Framingham, which contains a larger set of samples than Steno, is used to train CNN-based models. Finally, we performed transfer learning and fine-tuning using the pre-trained CNN on the Steno dataset to predict CVD risk. The CNN-based model with transfer learning achieved the highest AUCORC in Steno (0.855), outperforming ML models such as decision trees, K-nearest neighbors, least absolute shrinkage and selection operator (LASSO) support vector machine and TabPFN. This approach improved accuracy by 2% over the best-performing traditional model, TabPFN. To the best of our knowledge, this is the first study that evaluates the effectiveness of applying transfer learning and fine-tuning to tabular data using tabular-to-image approaches. Through the use of CNNs' predictive capabilities, our work also advances the diagnosis of CVD by providing a framework for early clinical intervention and decision-making support.

Noncommunicable diseases (NCDs) are the leading cause of morbidity and mortality worldwide. Cardiovascular diseases (CVDs) and diabetes are the most prevalent NCDs, causing 1.9 and 1.5 million deaths yearly. Individuals diagnosed with type 1 diabetes (T1D) are at high risk of developing CVDs. Machine learning (ML) models have provided outstanding results in different domains, including healthcare, allowing to obtain models with high predictive performance. The aim of this study was to develop an interpretable data-driven approach to predict the 10-year CVD risk for T1D older individuals, aiming to provide both reasonable predictive performance and the identification of risk factors associated with CVDs. Data from T1D individuals at the Steno Diabetes Center Copenhagen were used. Different ML-based models were considered, including KNN, decision tree, random forest, and multilayer perceptron (MLP). To enhance the predictive performance of ML models, the conditional tabular generative adversarial network (CTGAN) was used to create synthetic data and increase the size of the training data. Several filter and wrapper feature selection (FS) techniques were considered for identifying the most relevant features involved in CVD risk and enhancing the performance of the ML-based models used. To gain interpretability on predictive models, we used the post-hoc methods: SHAP and accumulated local effects. The experimental results showed a great performance of FS and ML-based models for predicting CVD risk. In particular, the MLP obtained the best results, with a mean absolute error of 0.0088 and mean relative absolute error of 0.0817. Regarding risk factors, age, Hba1c, and albuminuria were identified as crucial in CVD risk prediction, which is in line with recent clinical evidence. Our study contributes to identifying CVD risk and associated risk factors in a data-driven manner, helping to make early interventions and adequate treatments to prevent CVDs.

Background: The global incidence of skin cancers, particularly melanoma, the deadliest among them, has risen in recent decades. Artificial Intelligence (AI) has yielded remarkable results across diverse domains including healthcare. However, most AI models for skin cancer prediction primarily focus on a single data type or modality, such as images. Emerging research highlights the potential of multimodal approaches to enhance predictive performance and interpretability in clinical tasks. Methods: In this work, the effectiveness of multimodal fusion methods —including early and late fusion— for the detection of melanoma is assessed. Three public skin image datasets (ISIC, PH2, and Derm7pt) were considered, and two types of modalities were used, including dermoscopy images and tabular data comprising metadata, image features, and image embeddings. This study employed AI-based image classification and extended feature extraction to capture the color and texture characteristics of skin lesions, aiming to discover factors related to melanoma. Results: Experimental findings reveal that the proposed multimodal fusion methods outperform single-modality models. Early fusion methods achieve the highest classification results, with an overall AUCROC of 0.873, 0.963, and 0.857 for ISIC, PH2, and Derm7pt, respectively. Compared with models trained with single-modality data, fusion approaches exhibited notable improvements of 6.67% (ISIC), 8.27% (Derm7pt), and 7% (PH2) in AUCROC. Two post-hoc techniques were used for detecting salient regions in dermoscopy images and identifying the most relevant image features for melanoma diagnosis. Conclusions: These results hold promise for early skin lesion detection, contributing valuable support to dermatologists in their daily practice.

Artifacts in dermoscopic images---such as rulers, gel bubbles, or ink stains---can compromise automated skin lesion analysis and reduce diagnostic reliability. Despite their clinical relevance, prior research has primarily targeted the removal or impact assessment of single artifact types, without addressing their automatic detection or co-occurrence modeling.In this work, we introduce the first deep learning framework for multi-label artifact detection in dermoscopy, leveraging the Classification Transformer (C-Tran) to model dependencies between co-occurring labels through attention mechanisms. To address class imbalance and overconfident predictions, we propose the Class-Adaptive Focal loss with Confidence penalty (CAFC), which adaptively weights underrepresented classes and penalizes miscalibrated outputs. On the ISIC-2019 dataset, the model trained with CAFC surpasses both Binary Cross-Entropy and Focal losses, yielding the highest overall F1-score (0.87) and recall (0.91). The proposed framework also generalizes effectively to the external PH2 dataset, achieving an overall F1-score of 0.86 and a recall of 0.91, confirming robustness across domains and imaging distribution shifts. Additionally, the approach provides attention-based interpretability, revealing the spatial regions associated with each artifact type and enhancing transparency. Beyond its strong performance, the proposed approach establishes a foundation for future artifact-aware systems that could assist in image quality control, preprocessing, and bias auditing within clinical diagnostic pipelines.

Dermoscopic images have been the standard for evaluating skin lesions due to their high resolution and detail. However, advancements in digital photography and smartphone cameras have made it easier to capture highresolution images. Despite this, macroscopic images often suffer from uneven lighting, affecting the performance of deep learning models in tasks such as skin lesion segmentation. Recent supervised methods for low-light image enhancement (LLIE) require paired ground-truth images for training, limiting their applicability. To address this, we propose MacroE-Net, a modified version of the unsupervised zero-reference deep curve estimation model for LLIE on macroscopic images. MacroE-Net uses transfer learning and fine-tuning techniques and three no-reference loss functions, including the luminance ordinal distortion, multiscale image quality, and the integrated local non-linear image quality evaluator. We evaluated the effectiveness of enhanced images with MacroE-Net for skin lesion segmentation using macroscopic images from the Waterloo dataset and five U-Net-based architectures, including U-Net, Attention U-Net, Dense U-Net, Dense Attention U-Net, and U-Net 3+. MacroE-Net helped to improve segmentation performance, as demonstrated by both qualitative evaluation and quantitative metrics, including Dice Similarity Coefficient (DSC), Intersection-over-Union (IoU), and sensitivity. Dense Attention U-Net achieved the best segmentation metrics, with a DSC of 0.89, IoU of 0.80, and sensitivity of 0.89, outperforming results using original and state-of-the-art enhanced images by 2.9%, 3.3%, and 5.1%, respectively. By eliminating the need for paired datasets, our approach enhances flexibility and broadens its applicability. This highlights the potential of unsupervised LLIE methods to enhance the quality of macroscopic images and skin lesion segmentation.