Experience

Project: LCM-UNet — Lightweight Skin Lesion Segmentation using a Mamba-based U-Net

• Ultra-lightweight (~34K parameters, 0.06 GFLOPs) skin lesion segmentation network built on a Mamba-based state-space U-Net backbone, designed to detect melanoma-related lesion boundaries in dermoscopic images for deployment on low-cost, point-of-care hardware.
• Main contribution: the Residual Dilated Local Adapter (RDLA), a zero-initialized dilated convolution module that adds extra optimization capacity during training but is exactly fused back into the original kernel via structural reparameterization before inference, so deployment cost stays unchanged.
• Validated the approach with multi-seed statistical testing on ISIC 2017, ISIC 2018, and external PH2 datasets, showing consistent boundary-quality improvements over the baseline (image-level Wilcoxon significant at p < 1e-4).
• Promoted from Research Intern to Junior Research Assistant in recognition of research contributions.

• Independent research lab co-founded and led alongside collaborator Nishi Kanta Paul, building explainable, deployment-efficient AI for clinical use.
• Contributed to 6 peer-reviewed manuscripts spanning medical imaging (dermatology, ophthalmology, radiology) and wearable biosignal analysis (EEG, IMU): SleepEffFormer, WaveFoG, DR-LiteNet, AWEF-Net, PDGuard, and CoMAF-Polyp.
• Contributions: model design, cross-validation, statistical significance testing, and Grad-CAM-based explainability analysis — consistently achieving state-of-the-art parameter efficiency versus standard backbones (ResNet50, DenseNet201, etc.).

• Designed and developed digital solutions to support business operations, including website development, lead management automation, and workflow optimization.
• Responsible for maintaining web platforms, integrating business tools, and resolving technical issues to improve operational efficiency.