spatial-cell-composition

Submission for Global AI Hackathon'25 by Elucidata

Predicting Spatial Cell-Type Composition from Histology Images

This project tackles a real-world biomedical machine learning problem: predicting the spatial abundance of 35 cell types from high-resolution H&E-stained histology images using deep learning.

🧠 Motivation

Spatial transcriptomics is powerful but expensive. Histology is cheap but coarse. The challenge is to bridge the two with AI — mapping visual signals to molecular insights.

🧪 Dataset

• Provided via Kaggle (Elucidata Global AI Hackathon 2025)
• 6 training slides, 1 test slide
• For each slide:
  • HE image (float32 RGB)
  • ~2000 spatial transcriptomic spots
  • Each spot has abundance data for 35 cell types (C1–C35)

🔍 EDA Highlights

• Visualized tissue coverage and spot distribution
• Mapped cell-type heatmaps over tissue
• Found biologically meaningful spatial correlations
• Previewed high-abundance patches per cell type

🧠 Model

• ResNet18 pretrained on ImageNet
• Modified for 35-channel regression output
• Trained with MSE loss (can upgrade to Spearman rank loss)
• Extracted 224x224 patches per spot

🔄 Pipeline

1. Load .h5 data and extract patches
2. Train CNN on spot-wise cell composition
3. Predict on test slide S_7
4. Export submission.csv for Kaggle

📊 Results

• Baseline model achieved %
• Strongest correlations with structural cell types
• Future work: try EfficientNet, ViT, and integrate multimodal embeddings

🛠 Tech Stack

• Python, PyTorch, torchvision, h5py, pandas, matplotlib
• Jupyter Notebooks for reproducibility

📸 Sample Outputs

Include screenshots:

• Spot overlay on slide
• Heatmaps of cell type abundance
• Correlation matrix
• Top-k patch previews

📁 Submission

Includes a valid submission.csv for Kaggle.

---

Want to see the code? Check out:

• notebook/ — fully interactive Jupyter flows
• src/ — clean reusable modules