Transformer-Based Approach to Melanoma Detection.

Melanoma is a malignant cancer type which develops when DNA damage occurs (mainly due to environmental factors such as ultraviolet rays). Often, melanoma results in intense and aggressive cell growth that, if not caught in time, can bring one toward death. Thus, early identification at the initial stage is fundamental to stopping the spread of cancer. In this paper, a ViT-based architecture able to classify melanoma versus non-cancerous lesions is presented. The proposed predictive model is trained and tested on public skin cancer data from the ISIC challenge, and the obtained results are highly promising. Different classifier configurations are considered and analyzed in order to find the most discriminating one. The best one reached an accuracy of 0.948, sensitivity of 0.928, specificity of 0.967, and AUROC of 0.948.

Vision Transformer for femur fracture classification.

INTRODUCTION: In recent years, the scientific community focused on developing Computer-Aided Diagnosis (CAD) tools that could improve clinicians' bone fractures diagnosis, primarily based on Convolutional Neural Networks (CNNs). However, the discerning accuracy of fractures' subtypes was far from optimal. The aim of the study was 1) to evaluate a new CAD system based on Vision Transformers (ViT), a very recent and powerful deep learning technique, and 2) to assess whether clinicians' diagnostic accuracy could be improved using this system. MATERIALS AND METHODS: 4207 manually annotated images were used and distributed, by following the AO/OTA classification, in different fracture types. The ViT architecture was used and compared with a classic CNN and a multistage architecture composed of successive CNNs. To demonstrate the reliability of this approach, (1) the attention maps were used to visualize the most relevant areas of the images, (2) the performance of a generic CNN and ViT was compared through unsupervised learning techniques, and (3) 11 clinicians were asked to evaluate and classify 150 proximal femur fractures' images with and without the help of the ViT, then results were compared for potential improvement. RESULTS: The ViT was able to predict 83% of the test images correctly. Precision, recall and F1-score were 0.77 (CI 0.64-0.90), 0.76 (CI 0.62-0.91) and 0.77 (CI 0.64-0.89), respectively. The clinicians' diagnostic improvement was 29% (accuracy 97%; p 0.003) when supported by ViT's predictions, outperforming the algorithm alone. CONCLUSIONS: This paper showed the potential of Vision Transformers in bone fracture classification. For the first time, good results were obtained in sub-fractures classification, outperforming the state of the art. Accordingly, the assisted diagnosis yielded the best results, proving the effectiveness of collaborative work between neural networks and clinicians.

A Comparison of Deep Learning Techniques for Arterial Blood Pressure Prediction.

Continuous vital signal monitoring is becoming more relevant in preventing diseases that afflict a large part of the world's population; for this reason, healthcare equipment should be easy to wear and simple to use. Non-intrusive and non-invasive detection methods are a basic requirement for wearable medical devices, especially when these are used in sports applications or by the elderly for self-monitoring. Arterial blood pressure (ABP) is an essential physiological parameter for health monitoring. Most blood pressure measurement devices determine the systolic and diastolic arterial blood pressure through the inflation and the deflation of a cuff. This technique is uncomfortable for the user and may result in anxiety, and consequently affect the blood pressure and its measurement. The purpose of this paper is the continuous measurement of the ABP through a cuffless, non-intrusive approach. The approach of this paper is based on deep learning techniques where several neural networks are used to infer ABP, starting from photoplethysmogram (PPG) and electrocardiogram (ECG) signals. The ABP was predicted first by utilizing only PPG and then by using both PPG and ECG. Convolutional neural networks (ResNet and WaveNet) and recurrent neural networks (LSTM) were compared and analyzed for the regression task. Results show that the use of the ECG has resulted in improved performance for every proposed configuration. The best performing configuration was obtained with a ResNet followed by three LSTM layers: this led to a mean absolute error (MAE) of 4.118 mmHg on and 2.228 mmHg on systolic and diastolic blood pressures, respectively. The results comply with the American National Standards of the Association for the Advancement of Medical Instrumentation. ECG, PPG, and ABP measurements were extracted from the MIMIC database, which contains clinical signal data reflecting real measurements. The results were validated on a custom dataset created at Neuronica Lab, Politecnico di Torino.

The GH-EXIN neural network for hierarchical clustering.

Hierarchical clustering is an important tool for extracting information from data in a multi-resolution way. It is more meaningful if driven by data, as in the case of divisive algorithms, which split data until no more division is allowed. However, they have the drawback of the splitting threshold setting. The neural networks can address this problem, because they basically depend on data. The growing hierarchical GH-EXIN neural network builds a hierarchical tree in an incremental (data-driven architecture) and self-organized way. It is a top-down technique which defines the horizontal growth by means of an anisotropic region of influence, based on the novel idea of neighborhood convex hull. It also reallocates data and detects outliers by using a novel approach on all the leaves, simultaneously. Its complexity is estimated and an analysis of its user-dependent parameters is given. The advantages of the proposed approach, with regard to the best existing networks, are shown and analyzed, qualitatively and quantitatively, both in benchmark synthetic problems and in a real application (image recognition from video), in order to test the performance in building hierarchical trees. Furthermore, an important and very promising application of GH-EXIN in two-way hierarchical clustering, for the analysis of gene expression data in the study of the colorectal cancer is described.

The Growing Curvilinear Component Analysis (GCCA) neural network.

Big high dimensional data is becoming a challenging field of research. There exist a lot of techniques which infer information. However, because of the curse of dimensionality, a necessary step is the dimensionality reduction (DR) of the information. DR can be performed by linear and nonlinear algorithms. In general, linear algorithms are faster because of less computational burden. A related problem is dealing with time-varying high dimensional data, where the time dependence is due to nonstationary data distribution. Data stream algorithms are not able to project in lower dimensional spaces. Indeed, only linear projections, like principal component analysis (PCA), are used in real time while nonlinear techniques need the whole database (offline). The Growing Curvilinear Component Analysis (GCCA) neural network addresses this problem; it has a self-organized incremental architecture adapting to the changing data distribution and performs simultaneously the data quantization and projection by using CCA, a nonlinear distance-preserving reduction technique. This is achieved by introducing the idea of "seed", pair of neurons which colonize the input domain, and "bridge", a novel kind of edge in the manifold graph, which signals the data non-stationarity. Some artificial examples and a real application are given, with a comparison with other existing techniques.