Nine manuscripts across medical imaging, biosignal modelling, efficient deep learning, and trustworthy AI — developed at NIMISHES Lab and with collaborators at the Medical University of Vienna, Washington State University, and ELITE Research Lab. PDFs are linked where available.
Peer-reviewed and indexed on IEEE Xplore.
Int. Conf. on Advances in Computing, Communication, Electrical, and Smart Systems (iCACCESS), 2024 · IEEE Xplore
Transfer-learning approach to driver distraction recognition from in-vehicle video, evaluated on Bangladeshi road conditions.
Accepted for presentation at IEEE and international conferences in 2026.
5th IEEE Int. Conf. on Biomedical Engineering, Computer and Information Technology for Health (BECITHCON) · Dhaka, Bangladesh, 4–5 Sept. 2026
83.9% accuracy, 78.9% macro-F1 and Cohen’s κ = 0.765 on Sleep-EDF Expanded (78 recordings, subject-wise split), matching AttnSleep with 3–5× fewer parameters (~367K).
5th IEEE Int. Conf. on Biomedical Engineering, Computer and Information Technology for Health (BECITHCON) · Dhaka, Bangladesh, 4–5 Sept. 2026
F1 0.875 ± 0.017 and AUPRC 0.833 on the Kaggle TLVMC dataset under subject-independent ten-fold cross-validation — up to +5.4 pp F1 over CNN, Bi-LSTM, and Transformer baselines.
1st Int. Conf. on Next-Generation Electrical & Electronics, Computer Systems, and Technologies (iCONEECT) · Premier University, Chittagong, Bangladesh, 25–26 Sept. 2026
88.77% accuracy and QWK 0.884 on APTOS 2019 with only 4.69M parameters, outperforming ResNet50 (23.5M) while providing Grad-CAM explanations.
Submitted to IEEE conferences and peer-reviewed journals.
29th IEEE Int. Conf. on Computer and Information Technology (ICCIT), 2026
96.3% macro-F1 and 97.1% COVID-19 recall on the COVID-19 Radiography Database (21,165 images), a +1.4 pp gain over the strongest fixed-fusion baseline at under 0.6% extra parameters.
29th IEEE Int. Conf. on Computer and Information Technology (ICCIT), 2026
88.34% accuracy and pooled ROC-AUC 0.918 on the PADS cohort (469 participants) under subject-disjoint folds; 1.24M parameters at 18.3 ms per window on CPU.
IEEE Journal of Biomedical and Health Informatics (JBHI) · with ELITE Research Lab
Ultra-compact state-space segmentation at 34,180 parameters and 0.06 GFLOPs after reparameterization, improving IoU and DSC on ISIC 2017/2018 at identical inference cost.
IEEE Access · with Washington State University
Real-time behavioral anomaly scoring over blockchain evidence access with asynchronous TreeSHAP explanations: AUROC 0.953 on CERT r4.2 at 9.37 ms mean scoring overhead.
PLOS ONE
Dice 0.887 / IoU 0.839 at 35 FPS, with +15.7% Boundary-F1 over the strongest CNN baselines and external validation on CVC-ClinicDB, CVC-ColonDB, ETIS-Larib, and CVC-300.
Workshop manuscript in preparation. Not circulated publicly — please email me and the authors will gladly share a copy on request.
Target: WACV Workshop · with the Medical University of Vienna, Austria
Semi-supervised polyp segmentation that fuses pseudo-labels from a task-specific specialist segmenter and a vision foundation model, weighting each source by calibrated reliability rather than trusting either uniformly.