Keshav T.

Key Skills

Software

Other

Top Subject Matter

Brain MRI lesion detection and classification (FCD-II localization)

Self-supervised/representation learning for computer vision (AI training research)

Lung cancer (NSCLC subtype) self-supervised feature learning and fine-tuning

Top Data Types

Image

Top Task Types

Fine-tuning

Classification

Freelancer Overview

Research paper: AG-MIL (Annotation-Guided Multiple Instance Learning) for FCD-II Detection and Localisation in Brain MRI. Brings 2+ years of professional experience across complex professional workflows, research, and quality-focused execution. Core strengths include Other and PyTorch. Education includes Bachelor of Technology, National Institute of Technology Jamshedpur (2027) and Intermediate, Central Public School (2023). AI-training focus includes data types such as Medical, DICOM, and Image and labeling workflows including Fine-tuning, Classification, and Fine Tuning.

Labeling Experience

Research Contributor (Published Paper) - NIT Jamshedpur

ImageClassification

Contributed to research on AG-MIL, an annotation-guided multiple instance learning framework for FCD-II detection and localization in brain MRI. Proposed a weakly supervised learning strategy using MIL to model sparse lesion regions without dense annotations. Leveraged DenseNet-201 with channel attention to perform multi-scale feature extraction for clinical MRI tasks and efficient inference. • Designed a weakly supervised MIL framework for medical imaging • Modeled lesion regions using sparse supervision instead of dense labels • Implemented DenseNet-201 with channel attention for multi-scale features • Achieved high sensitivity and efficient inference on clinical datasets

2025 - 2025

Research paper: LungCLR — Self-Supervised Feature Learning for Lung Cancer Classification (Published 2025)

Fine-tuning

You developed a self-supervised representation learning pipeline using contrastive learning (SimCLR) for NSCLC subtype classification. You reduced labeled data dependency by applying EfficientNet-B3 fine-tuning to achieve very high accuracy with a small labeled sample size. The results were captured in a published 2025 conference paper titled LungCLR. • Implemented contrastive learning (SimCLR) to learn transferable representations. • Fine-tuned EfficientNet-B3 for NSCLC subtype classification with limited labeled samples. • Designed the training pipeline to maximize accuracy under data scarcity. • Prepared and published research findings at the 25th International Conference on BioInformatics and BioEngineering (2025).

2025 - 2025

Machine Vision and Intelligence Lab, NIT Jamshedpur — Research Intern (Apr 2025–Jun 2025)

OtherFine-tuning

During your 2025 research internship, you worked on self-supervised and representation learning methods for computer vision with an emphasis on contrastive learning and multi-scale feature extraction. You designed and evaluated deep learning models to improve generalization on real-world datasets, aligning with AI training workflows rather than manual annotation production. The internship focused on learning robust representations useful for downstream medical or vision tasks. • Applied contrastive learning concepts for self-supervised representation learning. • Built and evaluated deep learning models aimed at stronger generalization. • Focused on multi-scale feature extraction to improve learned representations. • Conducted model development and experimentation as part of AI training research.

2025 - 2025

Research paper: AG-MIL (Annotation-Guided Multiple Instance Learning) for FCD-II Detection and Localisation in Brain MRI (Published at DSPA, 2026)

OtherFine-tuning

You proposed a weakly supervised learning framework for brain MRI medical image classification using Multiple Instance Learning (MIL) to model sparse lesion regions without dense annotations. You leveraged multi-scale feature extraction with channel attention to improve sensitivity while keeping inference efficient on clinical MRI datasets. The work was presented as a published conference paper in 2026. • Defined a MIL-based training strategy for medical image classification and localization. • Designed feature extraction using DenseNet-201 with channel attention for multi-scale learning. • Evaluated performance on clinical MRI data emphasizing sensitivity. • Produced research outputs suitable for publication at DSPA 2026.

2026

Education

Work History