Oral Cancer Diagnosis with AI: Classifying tumor infiltrating lymphocytes in oral squamous cell carcinoma histopathology using joint learning framework

Oral squamous cell carcinoma (OSCC) is the most common oral cancer, with a poor prognosis and rising global incidence. Tumor-infiltrating lymphocytes (TILs) are crucial for OSCC grading, but current manual assessment methods are subjective and inconsistent. To solve this, we introduce OralTILs-ViT, a novel AI-powered framework that combines cellular and tissue-level insights for accurate and automated TILs classification. 

Key Innovations: 

• Joint Representation Learning: Unlike existing methods that focus only on tissue-level analysis, OralTILs-ViT captures both cellular and tissue interactions. It processes cellular density maps alongside H&E-stained tissue images, providing a more comprehensive understanding of TILs distribution. 

• TILSeg-MobileViT – A Novel Segmentation Model: Our first stage involves TILSeg-MobileViT, a multiclass segmentation model that utilizes a weakly supervised learning approach to generate precise cellular density maps for key components like tumor cells, stromal cells, and lymphocytes. This method significantly reduces the need for manual annotation, making TILs assessment more scalable and efficient. 

• Accurate and Automated TILs Classification: Using a dual-encoder model, OralTILs-ViT effectively classifies TILs infiltration levels into three clinically relevant categories: "Moderate to Marked," "Slight," and "None to Very Less", aligning with pathology practices. 

• Superior Performance and Reliability: Through extensive evaluations, our framework has been shown to outperform existing single-modality approaches. OralTILs-ViT achieved: 96.37% accuracy, 96.34% precision,  96.37% recall, & 96.35% F1 score. Additionally, TOPSIS analysis confirmed that our method ranks first across all TILs infiltration categories, reinforcing its effectiveness. 

• Advancing OSCC Diagnosis with AI: By combining AI and pathology, this approach enhances the reproducibility and objectivity of OSCC grading. The ability to automate TILs classification with high accuracy opens new doors for faster cancer diagnoses, better treatment decisions, and improved patient outcomes. 

This research represents a significant step forward in AI-driven cancer diagnostics, paving the way for more efficient, scalable, and precise pathology tools. 

Paper Link: https://www.nature.com/articles/s41598-025-86527-5 

Large Scale Summarization using Ensemble Prompts and In-context Learning Approaches 

The field of Information Assurance (IA) has evolved significantly over the years, shaped by technological advancements and emerging cybersecurity threats. To better understand this transformation, our latest research leverages Large Language Models (LLMs) and NLP techniques to analyze over 62,000 documents from 1967 to 2024, providing a decade-wise breakdown of key trends in IA research. 

What Makes Our Approach Unique?

• Innovative “Ensemble Prompts (Ev2)” Method: Our advanced prompt engineering technique enhances summarization by combining Chain of Density (CoD), Few-Shot Learning, role-based structuring, and adversarial testing to improve relevance, conciseness, and thematic richness.

• Comprehensive Literature Analysis: We use BERTopic for topic detection and LLMs for automatic summarization, creating a structured overview of IA research across decades.

• Beyond Traditional Summarization: Unlike existing methods that focus only on reducing redundancy, our approach provides targeted summaries for each decade, ensuring that key bibliographic references and logical topic progressions remain intact.

• Improved Performance: Outperforms traditional summarization methods by 16.7% to 29.6% in keyword definition tasks. Excels in 5 out of 7 tested metrics, while maintaining the logical integrity of references.

Paper Link: https://www.nature.com/articles/s41598-025-94551-8

Focal Cortical Dysplasia (Type II) Detection with Multi-modal MRI and a Deep-Learning Framework

Focal cortical dysplasia type II (FCD-II) is a prominent cortical development malformation associated with drug-resistant epileptic seizures that leads to lifelong cognitive impairment. 

• In this work, we proposed a deep learning framework to tackle the intricate task of identifying FCD-II and its subtypes (FCD-IIa and FCD-IIb). 

• The framework analyzes MRI images from multiple planes (axial, coronal, and sagittal) across T1-weighted and FLAIR MRI modalities, providing a comprehensive view of brain abnormalities.

• We presented approximately 700 result images and plots (main paper and supplementary results) from different detection tasks, which provide advanced and insightful details, highlighting significant findings from the analysis. https://doi.org/10.1038/s44303-024-00031-5

• We are shaping the future of epilepsy treatment, bringing us closer to a world where every patient can receive the right diagnosis and the best possible care. 

Paper Link: https://www.nature.com/articles/s44303-024-00031-5

Smart Fabric and E-Textile Pressure Sensor for High-Pressure Measurement

In this research a novel Smart Fabric Pressure Sensor (SFPS) was designed to measure high-pressure levels with remarkable precision and reliability. Through an iterative design process, the sensor’s range, sensitivity, and linearity were improved. A dedicated biasing circuit and a microcontroller-based data acquisition (DAQ) system were developed to ensure precise data collection.

Performance and Results:

The sensor was rigorously tested under controlled conditions, applying pressures between 4.63 and 300 kPa. 

• Linear measurement range: 4.63 to 74.08 kPa (r = -0.99)

• High sensitivity: 0.0296 V/kPa

• Accuracy: 97.72%

• Excellent repeatability: Standard error of the mean (SEM) = 0.25 kPa

These results validate the SFPS as a reliable, high-performance wearable sensor, paving the way for further research in smart fabrics and electronic textiles (e-textiles) for high-pressure monitoring applications.

Implications for Wearable Technology:

With its high accuracy and responsiveness, the SFPS could be a game-changer in fields such as:

• Healthcare – Monitoring pressure in prosthetics, compression therapy, and medical wearables

• Sports & Performance – Enhancing athlete monitoring and equipment optimization

• Industrial Safety – Detecting excessive pressure in protective gear

This research marks a significant step toward integrating e-textile-based sensors into real-world applications, making wearable technology more advanced and functional.

Paper Link: https://doi.org/10.1109/tim.2023.3312487 

Optimizing colorectal cancer segmentation with MobileViT-UNet and multi-criteria decision analysis

Colorectal cancer is one of the deadliest malignancies, and traditional manual examination methods often lack consistency and efficiency. To address this, our study explores advanced segmentation models—VGG16-UNet, ResNet50-UNet, MobileNet-UNet, and MobileViT-UNet. Notably, this is the first study to integrate 𝐌𝐨𝐛𝐢𝐥𝐞𝐕𝐢𝐓 as a UNet encoder.

Our findings highlight 𝐌𝐨𝐛𝐢𝐥𝐞𝐕𝐢𝐓-𝐔𝐍𝐞𝐭 𝐰𝐢𝐭𝐡 𝐃𝐢𝐜𝐞 𝐥𝐨𝐬𝐬 as the leading model, achieving impressive metrics:

•  Dice ratio: 0.944 ± 0.030

•  Jaccard index: 0.897 ± 0.049

•  Precision: 0.955 ± 0.046

•  Recall: 0.939 ± 0.038

Using 𝐓𝐎𝐏𝐒𝐈𝐒-𝐛𝐚𝐬𝐞𝐝 𝐦𝐮𝐥𝐭𝐢-𝐜𝐫𝐢𝐭𝐞𝐫𝐢𝐚 𝐝𝐞𝐜𝐢𝐬𝐢𝐨𝐧 𝐚𝐧𝐚𝐥𝐲𝐬𝐢𝐬, it ranked as the best model, significantly outperforming existing benchmarks in colorectal histopathology segmentation.

This breakthrough has the potential to improve the accuracy and efficiency of colorectal cancer diagnosis, making a real impact in healthcare.

Paper Link: https://peerj.com/articles/cs-2633/?td=wk

Epileptic Seizure Detection Using EEG and Machine Learning 

Epilepsy affects millions worldwide, and accurate, real-time detection is crucial for effective treatment. We propose a computationally efficient framework that integrates: 

• Hurst Exponent Analysis – Used to examine the long-term memory characteristics of EEG signals.

• Daubechies 4 Discrete Wavelet Transformation – Applied for feature extraction. 

• ANOVA & Random Forest Regression – Utilized for feature selection. 

• Machine Learning Models – Trained and tested multiple models, including: Random Forest Classifier,  Support Vector Machine (SVM), & Long Short-Term Memory (LSTM) Network.

Key Findings

• The Random Forest classifier outperforms other models, achieving: 97% accuracy &  97.2% sensitivity

•  Effective classification using single-channel EEG with minimal handcrafted features.

• Demonstrates strong generalization on the CHB-MIT scalp EEG database.

• Computationally efficient design, making it suitable for real-time implementation on edge hardware.

Paper Link: https://doi.org/10.3390/bioengineering12040355 

Development of Self-Powered Energy-Harvesting Electronic Module and Signal-Processing Framework for Wearable Healthcare Applications

We have developed a 𝐛𝐢𝐨𝐦𝐞𝐝𝐢𝐜𝐚𝐥 𝐰𝐞𝐚𝐫𝐚𝐛𝐥𝐞 𝐝𝐞𝐯𝐢𝐜𝐞 that advances the field of 𝐩𝐞𝐫𝐬𝐨𝐧𝐚𝐥𝐢𝐳𝐞𝐝 𝐡𝐞𝐚𝐥𝐭𝐡 𝐦𝐨𝐧𝐢𝐭𝐨𝐫𝐢𝐧𝐠 and 𝐞𝐚𝐫𝐥𝐲 𝐝𝐢𝐚𝐠𝐧𝐨𝐬𝐢𝐬.

In this work, we developed a 𝐬𝐞𝐥𝐟-𝐩𝐨𝐰𝐞𝐫𝐞𝐝 𝐩𝐨𝐫𝐭𝐚𝐛𝐥𝐞 𝐞𝐥𝐞𝐜𝐭𝐫𝐨𝐧𝐢𝐜 𝐦𝐨𝐝𝐮𝐥𝐞 designed for 𝐫𝐞𝐚𝐥-𝐭𝐢𝐦𝐞 𝐞𝐥𝐞𝐜𝐭𝐫𝐨𝐜𝐚𝐫𝐝𝐢𝐨𝐠𝐫𝐚𝐦 (𝐄𝐂𝐆) 𝐬𝐢𝐠𝐧𝐚𝐥 𝐩𝐫𝐨𝐜𝐞𝐬𝐬𝐢𝐧𝐠. Here's what makes it stand out:

• 𝐄𝐧𝐞𝐫𝐠𝐲-𝐄𝐟𝐟𝐢𝐜𝐢𝐞𝐧𝐭 𝐃𝐞𝐬𝐢𝐠𝐧: An onboard 𝐞𝐧𝐞𝐫𝐠𝐲-𝐡𝐚𝐫𝐯𝐞𝐬𝐭𝐢𝐧𝐠 system powered by lithium-ion batteries, silicon photodiode arrays, and solar panels ensures continuous operation.

• 𝐈𝐧𝐧𝐨𝐯𝐚𝐭𝐢𝐯𝐞 𝐒𝐢𝐠𝐧𝐚𝐥 𝐏𝐫𝐨𝐜𝐞𝐬𝐬𝐢𝐧𝐠: We introduced a 𝐝𝐲𝐧𝐚𝐦𝐢𝐜 𝐦𝐮𝐥𝐭𝐢-𝐥𝐞𝐯𝐞𝐥 𝐰𝐚𝐯𝐞𝐥𝐞𝐭 𝐩𝐚𝐜𝐤𝐞𝐭 𝐝𝐞𝐜𝐨𝐦𝐩𝐨𝐬𝐢𝐭𝐢𝐨𝐧 𝐟𝐫𝐚𝐦𝐞𝐰𝐨𝐫𝐤 framework to process ECG signals efficiently by removing overlapping samples.

• 𝐀𝐝𝐯𝐚𝐧𝐜𝐞𝐝 𝐌𝐚𝐜𝐡𝐢𝐧𝐞 𝐋𝐞𝐚𝐫𝐧𝐢𝐧𝐠: Our custom Random Forest with Deep Decision Tree (RFDDT) model achieved a remarkable 99.72% accuracy in offline ECG signal classification.

• 𝐒𝐮𝐬𝐭𝐚𝐢𝐧𝐚𝐛𝐥𝐞 𝐚𝐧𝐝 𝐏𝐨𝐫𝐭𝐚𝐛𝐥𝐞: With the integration of a BQ25505 IC, solar cells, and photodiodes, the module supports uninterrupted operation with ambient light.

The results demonstrate significant improvements in: Signal processing accuracy and efficiency, Energy optimization for wearable devices, and Low computing cost for personalized healthcare.

Paper Link: https://doi.org/10.3390/bioengineering11121252

Emotional, physiological, biochemical, and behavioral responses to acute stress and uncertainty in military personnel

Stress plays a critical role in perception, cognition, and decision-making, especially for individuals in high-stress professions like the military. This research study investigates how operationally relevant emotional stressors impact key cognitive functions, including:

• Memory Encoding & Recall – Does stress impair recognition memory?

• Spatial Learning & Navigation – How does stress influence orientation?

• Decision-Making Under Uncertainty – How do ambiguous threats affect rapid decisions?

Using a Threat of Shock (TOS) paradigm, we found:

• TOS increased sympathetic arousal but did not impact emotional or hormonal stress responses.

• Stress influenced decision times and confidence in a shoot/don’t shoot task, but not accuracy.

• Uncertainty (stimulus clarity) played a bigger role in decision-making than stress itself.

• Recognition memory and spatial orienting were unaffected by TOS.

Why It Matters: 

Understanding the cognitive effects of stress is crucial for military training, high-risk professions, and real-world decision-making. Our findings highlight the need for better stress-induction methods to simulate real-world military conditions in laboratory settings.

Paper Link: https://doi.org/10.1371/journal.pone.0312443

Trustworthy AI in Indian Healthcare

Our study critically reviews the state of AI adoption in healthcare and examines the challenges and opportunities of implementing 𝐭𝐫𝐮𝐬𝐭𝐰𝐨𝐫𝐭𝐡𝐲 𝐀𝐈 in the healthcare sector. Key takeaways from our research:

• 𝐀𝐈’𝐬 𝐏𝐨𝐭𝐞𝐧𝐭𝐢𝐚𝐥: AI-driven solutions can enhance disease detection, drug discovery, resource management, and diagnostic accuracy.

• 𝐓𝐞𝐜𝐡𝐧𝐨𝐥𝐨𝐠𝐲 𝐓𝐫𝐞𝐧𝐝𝐬: Most studies rely on machine learning (ML) and artificial neural networks (ANNs) for various healthcare applications.

• 𝐓𝐫𝐮𝐬𝐭 & 𝐄𝐭𝐡𝐢𝐜𝐬 𝐆𝐚𝐩: There is a significant gap in integrating ethical, legal, and regulatory-compliant AI in Indian healthcare settings.

• 𝐅𝐮𝐭𝐮𝐫𝐞 𝐑𝐨𝐚𝐝𝐦𝐚𝐩: The study highlights opportunities to develop AI solutions prioritizing patient safety, data privacy, and regulatory compliance. 

We hope this research sparks important conversations on building 𝐫𝐞𝐬𝐩𝐨𝐧𝐬𝐢𝐛𝐥𝐞 and 𝐢𝐦𝐩𝐚𝐜𝐭𝐟𝐮𝐥 𝐀𝐈 𝐬𝐨𝐥𝐮𝐭𝐢𝐨𝐧𝐬 for healthcare in India. Let's collaborate to make AI in healthcare 𝐭𝐫𝐮𝐬𝐭𝐰𝐨𝐫𝐭𝐡𝐲, 𝐬𝐜𝐚𝐥𝐚𝐛𝐥𝐞, 𝐚𝐧𝐝 𝐩𝐚𝐭𝐢𝐞𝐧𝐭-𝐜𝐞𝐧𝐭𝐫𝐢𝐜! 

Paper Link: https://doi.org/10.3390/ai6010010

Machine Learning Approach for Music Familiarity Classification with Single-Channel EEG   

Alzheimer’s patients temporarily recollected the forgotten memories after listening to familiar music.

• We predicted the familiarity of music using machine learning algorithms on data collected by single channel with an accuracy of 66.9%.

• This method is easy to use, noninvasive, and providing quick feedback for for tracking memory loss in dementia.

• This technology may potentially become become a therapeutic device that monitor and improve memory retention and communication.

• 
  

Paper Link: https://doi.org/10.1109/EMBC53108.2024.10782402 

Gaussian Distributed Semi-Analytic Reconstruction Method for Diffuse Optical Tomographic Measurement  

• Proposed a non-linear semi-analytic reconstruction method for continuous-wave diffuse optical tomography (CW-DOT), utilizing a Gaussian distribution framework.

• Experimental validation on phantom and in-vivo finger joint imaging of a healthy volunteer, showcasing the method's efficacy in providing consistently accurate estimates of tissue oxygenation levels (StO2) and oxygen capacity (CB), surpassing generalized methods that model photon paths as curved lines.

Paper Link: https://doi.org/10.1109/JSEN.2023.3326974 

Region-of-interest Diffuse Optical Tomography System 

We proposed a region-of-interest (ROI) tissue scanning method based on diffuse optical tomography (DOT) principles with a limited number of measurement data. Confining to ROI and limiting the measurement data can help us to rapidly survey the pathological and functional status of an interested region inside the tissue at a high speed. 

• We also have designed a wearable silicone rubber patch containing LEDs and photodetectors. The patch was mounted on the surface of the dynamic phantoms for carrying out the ROI experimental measurements. 

• The system can continuously acquire measurement data and reconstruct 3-D optical properties distributions of the dynamic resin and intralipid phantoms to visualize changes in optical properties. 

• Another imaging modality such as ultrasound (US) or magnetic resonance imaging (MRI) can be initially employed to get the structural information of the tissue, and then ROI DOT system can be explicitly used to monitor the interested region continuously. 

The results from this work further encouraged us to develop a functional ROI DOT system for 3-D spectroscopic imaging that derives the concentration of important chromophores such as oxyhemoglobin, deoxyhemoglobin, fat and water to characterize tumors located deep within the tissue.

Paper Link: https://doi.org/10.1063/1.4939054