Research | Mazharul Hossain

My Research

I develop machine learning methods for scientific images, with a focus on reliability when data or measurement conditions change.

Learning methods that account for shifts across sensors, scenes, and data distributions.

Reconstruction methods informed by how images are measured and formed.

Machine learning for scientific and biomedical imaging questions.

Problem: Rare anomalies and changing backgrounds make it hard for any single detector to work reliably across hyperspectral scenes and sensors.
Method: HUE-AD combined spectral-abundance estimates with three anomaly detectors by equal vote. GE-AD replaced fixed selection and weights with a learned ensemble; a later study compared supervised and unsupervised variants. GMoE-AD added neural experts and domain-adversarial training for cross-domain detection.
Evidence: GE-AD averaged 86.5% F1-macro on eight public images. GMoE-AD averaged 0.943 ROC-AUC across 22 images from seven datasets with multi-domain training and 0.910 in leave-one-dataset-out evaluation.

Problem: Reconstructing 3D structured-illumination microscopy images from noisy measurements requires both fine detail and fidelity to how the microscope forms an image.
Method: USR combines a learned image prior with repeated corrections based on the microscope point spread function. UPIGAN extends this physics-informed, unrolled approach with adversarial learning to improve reconstruction.
Evidence: Both studies used 18 fixed-cell mitochondrial samples. Their visual comparisons and reported PSNR and SSIM plots favor physics-informed unrolling over the respective non-physics baselines, but neither paper reports an independent external test.

Problem: Emergency responders need more than floor plans: they need to know where safety-critical objects are inside complex buildings.
Method: The Map901 team combined backpack LiDAR and 360° video, segmented safety objects in images, transferred their labels into 3D point clouds, and stitched and georeferenced the building scans.
Evidence: The studies surveyed seven Memphis buildings totaling 1.86 million square feet and addressed 30 classes of public-safety objects. The expanded study used 2,964 labeled images; its 3D maps still required manual cleanup of some labels.