Static Pitch
ChromaStrand
Author / Co-Author | University |
| Sadia Islam | Bangladesh Agricultural University |
Joyeeta Sarkar | Bangladesh Agricultural University |
Read More
DNA Sequence Comparison
Every living organism carries a molecular signature written in four letters. When two such signatures are placed side by side, their differences can reveal evolution, disease and function, yet the tools we use to read those differences have remained largely static for decades.
De Plot subverts this limitation by reimagining the dot matrix as a computational image substrate: a synthetic image generated from sequence data and subjected to a rigorous analytical pipeline. Comparing two DNA sequences is, at its core, like finding meaningful differences between two long paragraphs; De Plot makes those differences simultaneously visible and measurable.
The framework generates a binary dot plot from two input sequences, then cascades it through adaptive image processing: noise profiling, denoising, CLAHE enhancement, hybrid Otsu-local thresholding, morphological refinement and contour-based quantification. In parallel, it computes canonical genomic descriptors: GC content, Shannon entropy, Needleman-Wunsch alignment scores and percent identity. An interactive 3D double-helix interface with stage-synchronized bubble nodes unifies the analytical workflow into a coherent visual experience.
What renders De Plot paradigm-shifting is its syncretic architecture: transforming a traditionally static visualization artifact into a quantifiable analytical object by applying computer vision to bioinformatics in a manner that is both diagnostically rigorous and accessible to non-specialists. De Plot does not replace alignment methods; it extends them with image intelligence, offering a unified lens through which every molecular signature becomes simultaneously visible and measurable.
Team Inquisitive
Author / Co-Author | University |
Nadeef Muttaqin Chowdhury | BUET |
MD Rohan Sheikh | BUET |
Samiul Ahmed Pavel | BUET |
Test Case
In Bangladesh, rural and sub-urban patients face a problem of prolonged waiting time while taking treatment at distant city hospitals, which are often caused due to delay in processing of radiological images of MRI scans, CT scans etc. The waiting time is up to 148 minutes and 82 minutes at public and private hospitals respectively just for diagnostic tests.
This research introduces an automated screening framework to reduce the waiting time. It integrates two advanced Python computer vision models with machine learning analysis. Using brain tumors as a test case, the system utilizes an YOLOv8 model, a very fast model, which has been trained on 3064 MRI scans to instantly detect the presence and location of a tumor and for semantic segmentation of the tumor. It then deploys a ResUNet architecture to precisely outline the exact microscopic boundaries of the diseased tissue pixel-by-pixel.
Finally, by applying a digital bitwise masking algorithm, the system extracts only the overlapping area of the brain where both models agree. This two-way approach increases the accuracy of detection. Additionally, the patient's medical history is fed into a machine learning model that uses Scikit-learn to evaluate the overall gravity and immediate clinical risk of their condition.
Team Oblivion
Author / Co-Author | University |
Arnob Aich Anurag | AIUB |
Shamiul Islam | AIUB |
Md.Nasif Rafidi | AIUB |
Retinal disease classification from fundus photographs is difficult because diseases may have similar visual patterns, the images may not be of the same quality, and a class imbalance may dilute the errors of the minority class.
The research presented proposes a quality-aware framework for automating the screening for retinal diseases called ClaraVision-XAI, which takes an explainability-first approach. The verified local FD3611 split dataset has contained 3611 images distributed among five classes based on the folder from which they were derived.
The methodology applied in this research was to combine numerous retrained deep learning baselines, ClaraVision-style calibrated ensembles, GQA type architectural drawings, uncertainty analysis and XAI outputs. One representative C5 model from each of the following networks, ResNet, DenseNet, EfficientNet, MobileNet, Inception, Swin and ConvNeXt, was evaluated, with the best achieving performance of 88.87% accuracy, 81.18% macro precision, 80.38% macro recall, 80.58% macro f1 score, 93.44% AUC, and 4.30% ECE. This indicates that the model performed better on balance-based metrics and calibrated results than the original retraining family.
Therefore, this study finds that quality-aware calibrated AI is a promising area of research, however, further studies should include end-to-end training of the full hybrid model with larger clinically annotated data set.
CLIMATRIX
Author / Co-Author | University |
Pritha Priodorshini Audri | DU |
Musaffa Ahmed Elma | DU |
Shaira Evnat Ridita | DU |
The potential of Reflective Geotextiles in mitigating
glacier darkening and ice melt
Several studies document that glaciers are turning grey and are facing accelerated melting nowadays. Dust, black carbon, and soot particles create a layer which causes low albedo, i.e., less reflectivity and more absorption ability in the glaciers. As a result, the glacial ice melts at an abnormal rate, causing sea-level rise, floods, coastal erosion, and GLOFs.
Safeguarding glaciers is both an ecological prevention and a long-term investment in economic resilience. Previous research has stated that geotextiles can decrease ice melt and the reduction of emissions of dark carbon particles can reduce the deposition of pollutants in glaciers.
This study will evaluate whether investing in geotextiles will play an effective mitigation role in preventing glacier darkening. In this study, a comparison between the reduction in melt rate and albedo before and after the construction of geotextiles will be shown. Limitations of cost, environmental sustainability, and availability of biodegradable geotextiles are considered.
High-density polypropylene geotextiles (300–600 g/m²) reduce the ice melting by 69% under controlled field conditions. Biodegradable geotextiles are available but not widely, making it hard to use these at a large scale. Including material and labor cost, installing geotextiles will range from $1 to $3 dollar per square yard.
RegenSys
Author / Co-Author | University |
Muntasir Al Mamun | AUST |
Omeo Mahardad | AUST |
Sidratul Muntaha Sithi | AUST |
As processors grow more powerful in modern times, extreme heat and thermal throttling have become the critical bottlenecks. To solve this issue, we have developed RegenSys, a hybrid thermal management system that reimagines electronic cooling. Instead of the traditional continuous-flow liquid systems that dump heat into the environment which is inefficient, RegenSys utilizes a smart, ESP32-controlled batch fluid flow process. In this case, the coolant remains stationary at the heat source to maximize the heat absorption. Experimental results prove this thermal management system is more effective than the traditional one, actively dropping peak CPU temperatures from a critical temperature of 85°C down to a highly stable 55°C.
Crucially, RegenSys goes beyond just dissipating heat, it recycles it. By allowing the isolated fluid batches to reach thermal saturation, we preserve a steep 20°C temperature gradient. When displaced to a separate generation module, this trapped thermal energy interacts with thermoelectric generators to produce a peak open-circuit voltage of 1V.
While current thermoelectric material limits mean the system's overall power draw still exceeds its energy output, this proof of concept validates a powerful dual-purpose solution: actively protecting hardware from thermal degradation while successfully harvesting microelectronic waste heat.
Poster Presentation
Thesis Offended
Author / Co-Author | University |
Rahul Roy | BUET |
Musrat Binte Nur | BUET |
Dewan Nusaiba Ponkti | BUET |
Contact Author:
rahul.roy.90109@gmail.com
Hybrid Mesh Distribution Approach against
Traditional Hub and Spoke Model for the
Pharmaceutical Sector in Bangladesh during
Disruptions"
Team Thesis
Author / Co-Author | University |
Ariful Islam | AUST |
Md. Maruf | AUST |
Proactive Dengue Outbreak Management and
Medical Resource Allocation: A Case Study of
Dhaka South City Corporation, Bangladesh
Cardiac Implants via Machine Learning-Based
Alignment Detection and Thermal Prediction
Wireless power transfer (WPT) for cardiac implants eliminates battery replacement surgeries, but suffers efficiency loss from receiver coil misalignment. A 10 mm offset can reduce efficiency from 80% to 40%, generating excess tissue heating. Existing systems typically rely on reactive mechanisms, intervening only after thermal rise or charging instability has already occurred. We propose an ML system that performs external misalignment detection, adaptive impedance matching, and predictive thermal regulation, without FDA-flagged frequency changes.
The system operates at 300 kHz using resonant inductive coupling. External transmitter measures primary coil voltage and current to calculate input impedance
A 3-layer neural network using ReLU activation predicts optimal compensation capacitance from . Transmitter switches a binary-weighted capacitor bank to retune matching, restoring resonance and recovering efficiency.
A second neural network predicts tissue temperature 5 min ahead from implant thermistor readings, trained on Pennes' bioheat equation solutions.
Thus, by coupling adaptive resonance tuning with predictive thermal control, the proposed system achieves safer charging, reduced thermal risks, and improved power stability using existing implant sensor (thermistor) and external hardware (capacitor bank, impedance sensor, microcontroller).
Please Click Here to read More.
Adaptive Risk-Based Vaccine Allocation Decision Support System for Measles Outbreak Management in Bangladesh
In 2026 Bangladesh faced an upswing of measles, which recorded 72,070 suspected cases and 591 deaths across 61 of 64 districts, has exposed a fundamental weakness in conventional outbreak response: population-proportional vaccine allocation remains blind to where the disease is actively killing people. High-burden districts are chronically under-supplied while lower-risk areas absorb doses that could save more lives elsewhere (Jamil et al., 2026). To address this, the present study introduces ARVADS(Adaptive Risk-Based Vaccine Allocation Decision Support), a four-module operations research framework that converts real-time DGHS surveillance data into optimised allocation decisions within 48 hours. A Composite Risk Score Index was constructed for all 64 districts by integrating five epidemiological indicators, namely incidence rate, vaccination coverage gap, case fatality ratio, under-5 population proportion, and DGHS hotspot designation, using evidence-weighted coefficients and Min-Max normalisation. A Linear Programming model then minimised risk-weighted projected mortality under constrained supply scenarios of 50% and 75% with a 20% equity floor per district. The equity floor was incorporated to balance optimisation efficiency against distributional fairness(Crönert et al., 2024) . The framework is accompanied by an interactive web-based decision support dashboard that visualises real-time CRSI rankings for all 64 districts, renders a four-tier risk classification with colour-coded alert cards, presents grouped bar charts comparing LP-optimised versus population-proportional dose allocation, displays a lives-saved comparison panel, supports dynamic supply scenario toggling between 50% and 75% availability, and generates district-level field directives through an adaptive alert engine. The model identified Madaripur (CRSI: 48.58), Barguna (46.18), and Munshiganj (44.37) as the highest-priority districts and demonstrated a 5.0% improvement in deaths prevented over the proportional baseline. This deployable, data-driven framework ensures every available dose reaches where it will prevent the most deaths
Environmental Samples Using YOLOv8 Nano:
A Lightweight Approach for Edge Deployment
Microplastic contamination has become a rising environmental and public health crisis in Bangladesh. It affects not only rivers but also lakes, ponds, and even fish consumed daily. Studies show that Gulshan Lake contains up to 36 microplastics per liter of water. Microplastics have been found in human blood, lungs and brain, causing diseases including cancer. Detecting microplastics requires expensive laboratory equipment and expert knowledge, making it inaccessible to most students, researchers, and communities.
This project introduces an ML-powered mobile application that can detect microplastics in water samples by analyzing images. Users can upload photos through the app and the model identifies and highlights microplastic particles within seconds. The model architecture involves a Machine Learning model (YOLOv8) that can recognize patterns of microplastics from hundreds of sample images and deployment via Flutter. This enables detection accessible to anyone, providing instant detection using a smartphone and cheap TINYSCOPE microscope.
The system demonstrates strong and reliable detection performance, correctly identifying most microplastic particles while minimizing errors. It achieves 82% detection accuracy, with high precision in differentiating microplastics from non-plastic particles. While existing studies demonstrate ML detection accuracy in laboratory conditions, this work uniquely bridges the gap between model performance and community-level accessibility by delivering a field-deployable mobile application tailored to the Bangladesh freshwater context.
...