Papers & Publications

A Leak-Free Re-evaluation of Augmented Alzheimer's MRI with Deep Neural Networks

Hasanul Banna Himel · Akhlak Hossain · Suraiya Akhter · Mehreen Mallick Fiona · Md Shazzad Ahammed Shuvo · Nishat Samia

Re-examines reported gains from MRI data augmentation in Alzheimer's disease classification under a leak-free train/test protocol, isolating how much apparent performance depends on augmentation leakage across the split.

Token Smuggling and Log Injection in Open-Source AI Coding Assistants

Akhlak Hossain · Suraiya Akther Maisha · Rohit Ghosh · Sadia Islam Dia · Hasanul Banna Himel · Nishat Samia

Studies how open-source AI coding assistants handle untrusted input, focusing on token smuggling and log injection as attack surfaces that can bypass guardrails or corrupt audit trails.

mIoU Under Scrutiny: Benchmarking Self-Supervised Pretraining for Lung-Tumor CT Segmentation

Suraiya Akther Maisha · Akhlak Hossain · Mushfida Ferdous Maisha · Israt Jahan Rifa · Sadia Islam Dia · Hasanul Banna Himel · Nishat Samia

Benchmarks self-supervised pretraining strategies for lung-tumor CT segmentation and interrogates whether mean IoU alone is a reliable measure of segmentation quality in this setting.

PulseThread: A Verification-Centred Mobile Platform for Real-Time Blood Donor Mobilisation

Introduces PulseThread, a deployed mobile platform for coordinating real-time blood donor mobilisation in low-resource contexts like Bangladesh. Uses a verification-first design with single-use QR codes across three checkpoints (arrival, cross-match, donation) to prevent no-shows and fraudulent intermediaries.

Predicting Maternal Health Risk from Clinical Vital Signs Using Machine Learning

Develops a machine learning pipeline to classify maternal health risk (Low, Mid, High) using six routinely collected vital signs plus seven engineered hemodynamic features. An Extra Trees ensemble achieved 82.2% accuracy and 0.9396 AUC with SHAP explainability for clinical interpretability.