Hi, I'm Shubham GajjarBuilding practical AI for healthcare and real-world systems.

Abstract

Hybrid ResNet50–Vision Transformer model for seven-class skin lesion classification on HAM10000. Combines local feature learning with global context modeling, achieving 96.3% accuracy, macro F1 0.961, and AUC 1.00 across classes.

Key contributions

• •Combined frozen ResNet50 feature extractor with four‑head Vision Transformer blocks for balanced local and global feature learning.
• •Achieved macro F1 of 0.961 and AUC of 1.00 across all classes, presented to 100+ attendees.

Abstract

Attention-enhanced VGG16–MCA UNet for brain tumor segmentation in FLAIR MRI. Achieves 99.59% accuracy and 99.71% specificity on an LGG dataset, with ensemble variants improving Dice over standard UNet baselines.

Key contributions

• •Integrated multi‑channel attention with VGG16 encoder and UNet decoder for improved tumor delineation.
• •Designed full preprocessing pipeline with skull stripping, intensity normalization, and 256×256 resizing.

Abstract

Two-stage pipeline for hair-occluded skin lesion classification: U‑Net++ hair segmentation followed by an Extended ResNet50 classifier with Inverse Soft Mask Attention. Achieves 97.89% accuracy on HAM10000 with improved robustness across occluded/unoccluded regions.

Key contributions

• •Introduced Inverse Soft Mask Attention to handle hair‑occluded and unoccluded regions jointly.
• •Used dense residual blocks and Squeeze‑and‑Excitation modules with learnable feature aggregation.