Automated Bilateral Full-Femur Segmentation and Morphometric
Analysis from Multi-Resolution Clinical QCT Using Transfer Learning
and a Transformer Network
manuscript in preparation · B.Sc.
thesis, BUET · supervised by Dr. Mahmuda Naznin · defended Jun 2026
Trained and compared four architectures (3D-UNet, UNet, Swin-UNet,
UNETR) for 3D femur segmentation on 22 GB+ of clinical DICOM
volumes, reaching 97%+ accuracy; designed volumetric preprocessing,
a Dice + cross-entropy objective for severe class imbalance, and
morphological postprocessing for memory-constrained,
privacy-sensitive medical data.
Lightweight Knowledge Distillation of DNABERT-2 for Biological
Sequence Classification
preprint · Machine Learning Project ·
Jan–Apr 2026
Compressed DNABERT-2 (117M parameters) into a 26M-parameter student
retaining 93.3% accuracy against the 96.7% teacher, with a
4.5× inference speedup.
[preprint]
Shot-Boundary Detection for the IEEE Video & Image Processing
Cup
competition · IEEE VIP Cup · Jul 2025
Evaluated YOLOv8/v11 and TransNetV2 on a 2 GB+ video dataset,
improving shot-boundary detection from an 85% baseline to over 90%
accuracy.
