Projects

DermGemma

A multi-agent, multimodal RAG, AI dermatological assistant.

AI Research Open Source Agentic AI

GitHub Model Hub Docs

scAnalyzer

A Python toolkit for single-cell RNA sequencing (scRNA-seq) analysis.

Open Source

GitHub Model Hub Docs

BRCA Integrated Gradients

This 5-layer model predicts five clinical outcomes: histological, vital, and receptor status by concatenating 1,936 multi-omics features per patient node and using an LSTM unit to prevent GCN oversmoothing. Integrated Gradients provide explainability.

AI Research Open Source

GitHub Model Hub Docs

Single cell RNA sequencing

Single cell RNA sequencing using GenePT.

AI Research Open Source

GitHub Model Hub Docs

KEGG Pathway Enrichment Analysis

KEGG Pathway Enrichment Analysis using R and clusterProfiler.

AI Research Open Source

GitHub Model Hub Docs

Histopathologic CapsuleNet

Histopathologic CapsuleNet for cancer classification.

AI Research Open Source

GitHub Model Hub Docs

Genetic Algorithm Mona Lisa

This Python script implements a genetic algorithm that recreates a target image (the Mona Lisa) by iteratively evolving a population of 500 semi-transparent circles through processes of selection, crossover, and mutation to minimize the visual difference between the generated art and the original.

AI Research Open Source

GitHub Model Hub Docs

Multimodal Histopathology & Multi-omics Fusion

A multimodal model that combines histopathology image data and multi-omics data to classify breast cancer subtypes. It uses mamba, GNNs, and a pretrained ConvNext model.

AI Research Open Source

GitHub Model Hub Docs

HINN PyTorch implementation

PyTorch implementation of the paper: HINN: Hierarchical Input Neural Network identifies multi-omics biomarker for cognitive decline

AI Research Open Source

GitHub Model Hub Docs

IGCN Optimized

I improved the IGCN paper’s performance on the TCGA-BRCA dataset by replacing threshold-based adjacency with k-nearest neighbors (top-k cosine similarity) and implementing k-fold cross-validation. By streamlining the code, replacing cal_adj_mat_parameter calls with a single cosine_adj function call from utils.py, I increased accuracy to 0.808 (from 0.754), weighted F1 to 0.807 (from 0.767), macro F1 to 0.773 (from 0.734), and MCC to 0.719 (from 0.655). These optimizations also reduced runtime from 206 to 60 seconds, achieving a 3.4x speedup.

AI Research Open Source

GitHub Model Hub Docs

BRCA-multi-omics

A graph neural network (GNN) pipeline for binary survival classification.

AI Research Optimization Open Source

GitHub Model Hub Docs

ChestBLIP

Fine-tuned the BLIP model on MIMIC-CXR dataset.

AI Research Medical Imaging Optimization

GitHub Model Hub Docs

SegDICOM

A web-based application for manual segmentation of DICOM medical images. It allows users to upload DICOM files, navigate slice-by-slice, annotate regions of interest via a brush or erase tool, assign labels, and export segmentation maps in DICOM-SEG format. Annotations are also logged in a CSV for easy tracking and further analysis.

AI Tools Open Source

GitHub Docs

Spatial-attention ConvMixer

A new technique for gastrointestinal image classification.

AI Research Medical Imaging TensorFlow Open Source

GitHub Paper