The clinical adoption of deep learning in dermatology requires models that are not only highly accurate but also transparent and trustworthy. To address this dual challenge, this study presents a systematic investigation into deep feature fusion, exploring how to effectively combine complementary representations from diverse neural network architectures. We design and rigorously evaluate six distinct fusion models, first investigating depth-wise and channel-wise strategies for integrating features from powerful Convolutional Neural Network (CNN) backbones, and subsequently incorporating the global contextual awareness of Vision Transformers (ViTs). Evaluated on the challenging 7-class HAM10000 dataset, our optimized architecture achieves a weighted average Precision, Recall, and F1 score of 90%, demonstrating superior diagnostic performance. Crucially, our comprehensive explainable AI (XAI) analysis using Grad-CAM and SHAP reveals that the fusion strategy directly dictates the model’s clinical interpretability; our most effective models learn to base their predictions on salient dermatological features, such as border irregularity and color variegation, in a manner that aligns with expert reasoning. This work provides a robust framework and valuable architectural insights for developing the next generation of high-performing, clinically reliable, and transparent AI-powered diagnostic tools.

Melanoma, recognized as the most lethal form of cancer, originates from the uncontrolled proliferation of melanocytes, cells responsible for melanin production. Identifying melanoma lesions poses a formidable challenge due to their visual resemblance to non-cancerous lesions. This paper presents an end-to-end skin lesion prediction framework that includes data collection, preprocessing, data augmentation, the state-of-the-art YOLOv8 prediction model, and evaluation procedure. To train and assess the proposed framework, two benchmark medical datasets were collected and used: (1) PH2 and (2) ISIC2018. The preprocessing procedure is conducted to improve prediction performance, and the augmentation process is used to increase the training set size to meet the requirements of the AI model. Subsequently, the state-of-the-art deep learning YOLOv8 model is adopted and fine-tuned as the core prediction component for the proposed framework. The detection results achieved an overall mAPs@0.5 of 99.50% and 97.30% using the PH2 and ISIC2018 datasets, respectively. These promising and encouraging evaluation results appear to be practically useful for establishing an end-to-end prediction framework for skin lesions.

Global indoor localization algorithms enable the robot to estimate its pose in pre-mapped environments using sensor measurements when its initial pose is unknown. The conventional Adaptive Monte Carlo Localization (AMCL) is a highly efficient localization algorithm that can successfully cope with global uncertainty. Since the global localization problem is paramount in mobile robots, we propose a novel approach that can significantly reduce the amount of time it takes for the algorithm to converge to true pose. Given the map and initial scan data, the proposed algorithm detects regions with high likelihood based on the observation model. As a result, the suggested sample distribution will expedite the process of localization. In this study, we also present an effective resampling strategy to deal with the kidnapped robot problem that enables the robot to recover quickly when the sample weights drop-down due to unmapped dynamic obstacles within the sensor’s field of view. The proposed approach distributes the random samples within a circular region centered around the robot’s pose by taking into account the prior knowledge about the most recent successful pose estimation. Since the samples are distributed over the region with high probabilities, it will take less time for the samples to converge to the actual pose. The percentage of improvement for the small sample set (500 samples) exceeded 90% over the large maps and played a big role in reducing computational resources. In general, the results demonstrate the localization efficacy of the proposed scheme, even with small sample sets. Consequently, the proposed scheme significantly increases the real-time performance of the algorithm by 85.12% on average in terms of decreasing the computational cost.