Acropolis Institute of Technology and Research

Nowadays, skin cancer is one of the most common cancers, and early diagnosis has a significant influence on improving the outcome of the patient. In this paper, a novel hybrid Deep Learning (DL) based approach is presented for the classification of skin lesions where both Convolutional Neural Networks (CNNs) as well as architectures-based on Transformer are implemented together. As such, the primary goal of the proposed model would be to enhance both the effectiveness and efficiency of the diagnosis of skin lesions by focusing on spatial as well as contextual features of the images. The CNN layer extracts rich features from the input images of skin lesions, while the Transformer layer captures the longer dependencies and contextual information within the images, which consequently allows for better distinguishing between benign and malignant lesions. Moreover, the model is further enhanced with clinical metadata pertinent to patient information to increase the accuracy of the diagnostic process. To examine the efficacy of the proposed approach, an extensive experiment was conducted using a publicly available skin lesion dataset (HAM10000-2018, ISIC-2019, and DermNet-2019); further comparisons were done with traditional CNN-based classifiers. The results attained by hybrid model overperform traditional classification performance in terms of diagnosis and efficiency. Advancement in the automation of medical image analysis provides us with benefits and strengths of using both CNNs and Transformers in the classification of skin lesions.

This study demonstrates a large-area, high-performance RF T-gate GaN HEMT fabricated using an 8-inch CMOS back-end-of-line (BEOL)-compatible process. This compatibility enables the efficient integration of GaN-based RF components with CMOS circuits, thereby enhancing overall functionality for next-generation applications in advanced electronics and packaging. To improve the DC and RF performance of the RF device, supercritical fluid nitrogen (SCFN) treatment was applied and optimized at 250 degrees C for 30 min, effectively reducing trap states on both the AlGaN surface and the AlGaN/GaN interface. Under the optimized condition, the SCFN-treated device exhibited a 21 % increase in maximum drain current (IDSon,max) by 21 %, a 13 % increase in maximum transconductance (gm,max), a 59 % reduction in subthreshold swing (SS), and a 40 % reduction in on-resistance (Ron). The device also achieved a notably low drain-induced barrier lowering (DIBL) of 92 mV/V, compared with 175 mV/V for the untreated device. Furthermore, RF performance was substantially improved, with the SCFN-treated device achieving fT/fmax values of 60/101 GHz-outperforming the 49/72 GHz in untreated devices, at a T-gate length of 0.18 mu m. To further clarify these enhancements, a TCAD simulation was conducted to analyze electron trapping in the drift region (gate-drain access region) and its impact on band bending and partial depletion of the two-dimensional electron gas (2DEG). These results provide a comprehensive understanding of the mechanisms driving the observed performance improvement.

Pneumonia screening in an environment with limited capacity for diagnostic imaging modalities can utilize predictive models trained on routine clinical data including symptoms, vital signs, and laboratory tests. This paper evaluated the performance of the Logistic Regression, Decision Tree, Random Forest, Support Vector Machine, and Extreme Gradient Boosting classifiers on a public pediatric tabular database with 710 instances. The tabular database was preprocessed by dropping the X-ray view since it was not fully developed for the purpose of this analysis. The categorical values were filled with the most frequent ones and then one-hot encoded, while the numerical values were replaced with their median values and standardized within the same transformation step. The resulting 710 instances were stratified-split into training (80%) and testing (20%) samples, of which 142 were used to score the algorithm performance. Three of the five machine learning algorithms demonstrated a perfect precision (98.59%) and an excellent F1-score (98.77%). XGBoost showed a perfect ROC-AUC (.9999). Overall, the Decision Tree was selected as the final interpreter model since it matched the top performer’s scores while being relatively easily interpretable. This analysis should be considered a proof of concept due to the possibility of the database being artificially generated or containing anonymous data. Only one holdout sample was used to evaluate the generalizability of the training set performance, and there were no external validation sites involved. Further investigation with more extensive multi-center data and more rigorous validation, calibration, and clinical outcomes validation procedures is needed before such a model can be subject to real-world implementation.

The quick spread of fake news on digital platforms is a big threat to public trust, democratic processes, and the health of society as a whole. Keyword-based filtering and classical machine learning are examples of traditional detection methods that have not been able to handle the complexity and subtlety of modern misinformation. This paper suggests a better way to find fake news using BERT (Bidirectional Encoder Representations from Transformers), a deep learning model based on Transformers that can accurately capture the subtleties of language and the meaning of words in context. We fine-tune the model on benchmark datasets like LIAR, FakeNewsNet, and ISOT, and then we test it using metrics like precision, recall, F1-score, accuracy, and ROC-AUC. The BERT-based model does much better than traditional classifiers like Logistic Regression, SVM, CNN, and LSTM, with an accuracy of 92.4% and a ROC-AUC of 95.1%. This study goes into more detail about problems like computational complexity and changing patterns of false information. It also looks at possible future directions, such as using lightweight models, multimodal analysis, and explainable AI. The results show that BERT could be a good and scalable way to fight fake news in real life.

Effective alumni management is essential for maintaining lifelong connections and fostering institutional growth. However, most educational institutions lack a centralized system to manage alumni data, resulting in scattered records, limited communication, and missed opportunities for collaboration. ReConnect is a web-based platform developed to streamline alumni data management and enhance engagement between alumni, students, and institutions. The system provides features such as alumni registration, event management, mentorship programs, and career opportunities within a single, secure interface. It is designed using the MERN stack and enforces data accuracy and privacy through role-based access control.