Deep Learning has contributed significantly to the advances made in the fields of Medical Imaging and Computer Aided Diagnosis (CAD). Although a variety of Deep Learning (DL) models exist for the purposes of image classification in the medical domain, more analysis needs to be conducted on their decision-making processes. For this reason, several novel Explainable AI (XAI) techniques have been proposed in recent years to better understand DL models. Currently, medical professionals rely on visual inspections to diagnose potential diseases in endoscopic imaging in the preliminary stages. However, we believe that the use of automated systems can enhance both the efficiency for such diagnoses. The aim of this study is to increase the reliability of model predictions within the field of endoscopic imaging by implementing several transfer learning models on a balanced subset of Kvasir-capsule, a Wireless Capsule Endoscopy imaging dataset. This subset includes the top 9 classes of the dataset for training and testing. The results obtained were an F1-score of 97%±1% for the Vision Transformer model, although other models such as MobileNetv3Large and ResNet152v2 were also able to achieve F1-scores of over 90%. These are currently the highest-reported metrics on this data, improving upon prior studies done on the same dataset. The heatmaps of several XAI techniques, including GradCAM, GradCAM++, LayersCAM, LIME, and SHAP have been presented in image form and evaluated according to their highlighted regions of importance. This is in an effort to better understand the decisions of the top-performing DL models and look beyond their black-box nature.

Variable (feature) selection plays an important role in data analysis and mathematical modeling. This paper aims to address the significant lack of formal evaluation benchmarks for feature selection algorithms (FSAs). To evaluate FSAs effectively, controlled environments are required, and the use of synthetic datasets offers significant advantages. We introduce a set of ten synthetically generated datasets with known relevance, redundancy, and irrelevance of features, derived from various mathematical, logical, and geometric sources. Additionally, eight FSAs are evaluated on these datasets based on their relevance and novelty. The paper first introduces the datasets and then provides a comprehensive experimental analysis of the performance of the selected FSAs on these datasets including testing the FSAs’ resilience on two types of induced data noise. The analysis has guided the grouping of the generated datasets into four groups of data complexity. Lastly, we provide public access to the generated datasets to facilitate bench-marking of new feature selection algorithms in the field via our Github repository. The contributions of this paper aim to foster the development of novel feature selection algorithms and advance their study.

This paper aims to explore the possibility of utilizing vision transformers (ViTs) for on-edge medical diagnostics by experimenting with the Kvasir-Capsule image classification dataset, a large-scale image dataset of gastrointestinal diseases. Quantization techniques made available through TensorFlow Lite (TFLite), including post-training float-16 (F16) quantization and quantization-aware training (QAT), are applied to achieve reductions in model size, without compromising performance. The seven ViT models selected for this study are EfficientFormerV2S2, EfficientViT_B0, EfficientViT_M4, MobileViT_V2_050, MobileViT_V2_100, MobileViT_V2_175, and RepViT_M11. Three metrics are considered when analyzing a model: (i) F1-score, (ii) model size, and (iii) performance-to-size ratio, where performance is the F1-score and size is the model size in megabytes (MB). In terms of F1-score, we show that MobileViT_V2_175 with F16 quantization outperforms all other models with an F1-score of 0.9534. On the other hand, MobileViT_V2_050 trained using QAT was scaled down to a model size of 1.70 MB, making it the smallest model amongst the variations this paper examined. MobileViT_V2_050 also achieved the highest performance-to-size ratio of 41.25. Despite preferring smaller models for latency and memory concerns, medical diagnostics cannot afford poor-performing models. We conclude that MobileViT_V2_175 with F16 quantization is our best-performing model, with a small size of 27.47 MB, providing a benchmark for lightweight models on the Kvasir-Capsule dataset.

Wideband spectrum sensing demands ultra-low latency and high accuracy to detect spectrum holes, yet deploying deep learning (DL)-based models on resource-constrained edge devices is challenging due to high computational costs. This letter proposes quantization-aware training (QAT) to optimize DL-based spectrum sensing models for low-power, low-memory deployment with fast inference. Using a hardware-oriented approach and data-driven quantization scaling, the models retain near-identical performance across varying signal-to-noise ratio (SNR) levels. Real-time deployment on the Sony Spresense shows 72% model size reduction, 51% faster inference, and 7% lower power consumption, confirming the feasibility of QAT-optimized models for spectrum sensing on the edge.

Vision Transformers (ViTs) have emerged in recent years as a powerful architecture for image classification tasks and beyond. However, due to their back-box nature, our understanding of their learning dynamics is minimal. In this work, we study ViT learning dynamics through the lens of class separation, tracking inter- and intra-class structure, propagating from the patch-embedding layer to the classifier head. Across four datasets and extensive hyper-parameter configurations, we characterize two primary regularities. Firstly, we observe a layer-wise consistency, where separation fuzziness spikes between the input projection and the first transformer block, then decays exponentially with depth. Secondly, provided that the first law holds, fuzziness at the final layer also decays exponentially as a function of the training epochs. The persistence of these laws across datasets and hyper-parameters suggests an intrinsic relationship to ViT training. Our study supplies a quantitative foundation for interpreting the design of transformer-based vision models. Code and reproducibility scripts are available through https://github.com/DaraVaram/vit-data-separation.