Publications

Authors: Nishi Kanta Paul, Md Shihabul Islam Shovo, Israt Jerin Esha, Adrita Rahman

5th IEEE Int. Conf. on Biomedical Engineering, Computer and Information Technology for Health (BECITHCON 2026), Dhaka, Bangladesh, 4–5 September 2026

A lightweight CNN-Transformer for automated sleep staging from a single EEG channel, paired with a non-parametric Transition-Aware Smoothing (TAS) layer that suppresses physiologically implausible stage transitions.

• 83.9% accuracy, 78.9% macro-F1, Cohen's kappa 0.765 on Sleep-EDF Expanded (78 all-night recordings, subject-wise split)
• Matches AttnSleep with 3–5x fewer parameters (~367K)
• Transformer encoder adds +6.6pp macro-F1; TAS adds +1.8pp with zero trainable parameters
• Attention maps recover physiologically meaningful sleep-stage EEG signatures

First author. Authors: Md Shihabul Islam Shovo, Nishi Kanta Paul, Adrita Rahman, Israt Jerin Esha

5th IEEE BECITHCON 2026, Dhaka, Bangladesh, 4–5 September 2026

A dual-branch CNN-Transformer that fuses discrete wavelet sub-band descriptors through a sigmoid gating layer to detect freezing-of-gait episodes from wrist-worn accelerometers under severe class imbalance.

• F1 0.875 ± 0.017 and AUPRC 0.833 ± 0.020 on the Kaggle TLVMC FoG dataset
• Subject-independent grouped ten-fold cross-validation
• Up to +5.4pp F1 over 1D CNN, Bi-LSTM, and vanilla Transformer baselines
• Temporal saliency maps align with reported FoG onset characteristics

Authors: NIMISHES Lab (incl. Md Shihabul Islam Shovo)

1st Int. Conf. on Next-Generation Electrical & Electronics, Computer Systems, and Technologies (iCONEECT 2026)

A dual-branch network that fuses 1,280-d EfficientNetB0 features with a 52-d handcrafted lesion descriptor (vessel density, exudate intensity, texture entropy, HSV histograms), then applies ADASYN in feature space to correct five-grade class imbalance without touching the images.

• 88.77% ± 0.61 accuracy, QWK 0.884 ± 0.010, macro-F1 0.792 ± 0.008 on APTOS 2019
• Only 4.69M parameters — beats ResNet50 (23.5M) by ΔQWK = 0.022
• Two-phase training: balanced head warm-up, then end-to-end fine-tuning
• Grad-CAM saliency maps for post-hoc clinical explainability

Authors: NIMISHES Lab (incl. Md Shihabul Islam Shovo)

Replaces fixed late-fusion rules with a lightweight two-layer MLP attention gate that produces input-adaptive weights for ResNet50, InceptionV3, and DenseNet201, resolving the inter-model feature conflict where one backbone dominates the fused decision.

• 96.3% macro-F1 and 97.1% COVID-19 recall on the COVID-19 Radiography Database (21,165 images, 4 classes)
• +1.4pp macro-F1 and +1.7pp COVID recall over the strongest fixed-fusion baseline
• Gating adds under 0.6% of total model parameters
• Better cross-dataset generalisation on an external CXR corpus

Authors: NIMISHES Lab (incl. Md Shihabul Islam Shovo)

A compact 1D-convolution, bidirectional LSTM, and multi-head self-attention network that classifies raw wrist-worn accelerometer and gyroscope streams, targeting frequent monitoring outside the clinic.

• 88.34% ± 1.01 accuracy, 83.79% balanced accuracy, pooled held-out ROC-AUC 0.918 on the PADS cohort (469 participants)
• 77.34% accuracy on the harder PD vs. differential-diagnosis task
• Strictly subject-disjoint five-fold cross-validation
• 1.24M parameters, 18.3 ms per 4.8s window on CPU — smartphone-class deployment

First author. Authors: Md Shihabul Islam Shovo, Nishi Kanta Paul, Kishor Morol

Target: journal manuscript, target submission September 2026

Strengthens the local pathway of an ultra-compact state-space segmentation network with a zero-initialized dilated depthwise adapter (RDLA) that merges exactly into a single kernel after training — more capacity during learning, an unchanged graph at inference.

• 34,180 parameters and 0.05974 GFLOPs after reparameterization — 4 parameters more than the reference, identical cost
• +0.48 / +0.37pp IoU and +0.29 / +0.22pp DSC over a retrained reference backbone on ISIC 2017 and ISIC 2018
• External PH2 transfer: 86.37% ± 0.29 IoU vs. 85.91% ± 0.29 for the reference
• Fused and unfused models emit identical masks to numerical tolerance; image-level Wilcoxon significant at p < 1e-4

Authors: Nishi Kanta Paul, Md Shihabul Islam Shovo, Camila Gonzalez (Medical University of Vienna)

Target: WACV 2027 Workshop (Computer Vision Foundation)

A calibration-aware adaptive fusion framework ("Consensus of Multi-source Adaptive Fusion") for pseudo-label generation in semi-supervised polyp segmentation from colonoscopy. Fuses candidates from a domain-adapted EMA specialist and a frozen SAM-Med2D foundation model, gated by a 5-descriptor reliability score via a calibrated 1.4K-parameter MLP gate.

• 88.46% Dice on Kvasir-SEG with 10% labels; fused pseudo-labels recover 98.6% of a per-image oracle
• +2.44pp average Dice over the best baseline across four unseen datasets under domain shift
• Calibration reduces expected calibration error (ECE) from 0.147 to 0.038
• All added machinery is training-time only; deployed model runs at 38.2 FPS