Latchan Chhetri

Latchan Chhetri

he/him

AI Researcher | B.Tech CSE (AI & DS) @ MIT, Sikkim | Research Intern @ ISI Kolkata & Univ. of Salford.

Publications

Risk-Controlled Urban Change Detection: Conformal Prediction Wrappers for Provable Reliability in High-Resolution Satellite Imagery

A rigorous statistical approach applying Marginal Split Conformal Prediction to transformer-based satellite change detection, achieving distribution-free coverage bounds and SOTA F1 scores.

Recommended citation: A. Kumar, A. Mukherjee, H. Das, L. Chhetri, P. Ghosal. (2026). "Risk-Controlled Urban Change Detection: Conformal Prediction Wrappers for Provable Reliability in High-Resolution Satellite Imagery." ICCI (Under Review).

Intrinsic Neural Firewalls for Cyber-Physical Systems: Robust Anomaly Rejection via Deep Delta Residual Overwrites

An intrinsic defense architecture (DDL-FW) for Cyber-Physical Systems that embeds anomaly detection directly into residual connections, achieving an F1 score of 0.7206 on the HAI 21.03 benchmark.

Recommended citation: R. Das, L. Chhetri, A. Kumar, P. Ghosal. (2026). "Intrinsic Neural Firewalls for Cyber-Physical Systems: Robust Anomaly Rejection via Deep Delta Residual Overwrites." WIN 6.0 (Under Review).

Beyond Limited Labels: Safe Semi-Supervised Learning for Malaria Diagnosis

An uncertainty-guided semi-supervised learning framework that integrates Monte Carlo dropout-based epistemic uncertainty estimation to reduce silent failure rates by 29.2% in resource-constrained environments.

Recommended citation: L. Chhetri, Aman Kumar. (2026). "Beyond Limited Labels: Safe Semi-Supervised Learning for Malaria Diagnosis." IEEE DSAA (Under Review).

Optimizing Deep Learning for Brain Tumor Classification: A Comparative Ablation Study of Preprocessing and Augmentation Strategies

A comprehensive ablation study isolating the impact of specific data preprocessing and augmentation pipelines, proposing a novel “destructive vs. constructive” taxonomy for medical image processing.

Recommended citation: L. Chhetri, A. Kumar. (2026). "Optimizing Deep Learning for Brain Tumor Classification: A Comparative Ablation Study of Preprocessing and Augmentation Strategies." IEEE GCON.

Deep Delta Vision Mamba: A Lightweight State Space Architecture with Deep Delta Learning for Efficient Remote Sensing

A hierarchical State Space Model (SSM) architecture integrating 2D Deep Delta Learning, achieving 510 FPS and outperforming ResNet50 by 13.1x in deployment efficiency for satellite edge computing.

Recommended citation: L. Chhetri, A. Kumar, G. Sarma. (2026). "Deep Delta Vision Mamba: A Lightweight State Space Architecture with Deep Delta Learning for Efficient Remote Sensing." IEEE CONECCT.

Ionospheric TEC Forecasting via Deep Delta Learning and Conformal Prediction

A CNN-DDL architecture featuring a dynamic beta-gate and Marginal Split Conformal Prediction for real-time solar wind modeling, achieving zero-shot cross-solar-cycle generalization.

Recommended citation: A. Kumar, L. Chhetri, P. Ghosal. (2026). "Zero-Shot Cross-Solar Cycle Generalization of Ionospheric Total Electron Content Dynamics using a Deep Delta Learning Convolutional Neural Network." Advances in Space Research (Under Review).

Geometric Residual Learning for 3D Medical Image Segmentation via DDL-Mamba

A novel Delta-Mamba framework integrating Deep Delta Learning blocks into a 3D U-Mamba backbone, featuring Softplus-constrained weights to stabilize representation drift in dense 3D State Space Models.

Recommended citation: L. Chhetri, et al. (2026). "Geometric Residual Learning for 3D Medical Image Segmentation via DDL-Mamba." In Progress.

AHF-RBFNet: Attention-Guided Hierarchical Fusion U-Net with Learnable Fuzzy Membership Kernels for Medical Image Segmentation

Integrating a learnable Radial Basis Function (RBF) fuzzy membership kernel into a hierarchical U-Net to dynamically filter semantic uncertainty, achieving SOTA IoU on skin lesion datasets.

Recommended citation: L. Chhetri, et al. (2026). "AHF-RBFNet: Attention-Guided Hierarchical Fusion U-Net with Learnable Fuzzy Membership Kernels for Medical Image Segmentation." In Progress.