Glioma Brain‑Tumor Cell Classification

Timeline: Aug 2024 – Nov 2024   |   Collaborators: IIT Jodhpur × AIIMS Jodhpur   |   Live demo: https://glioma‑cells‑demo.streamlit.app   |   Repo: https://github.com/YOUR‑ORG/glioma‑cell‑classify

---

1  Project at a Glance

Automatically locating and classifying the cellular landscape of glioma biopsy images accelerates diagnosis, treatment planning, and therapeutic research. We benchmarked three deep‑learning pipelines built around YOLOv8 detection and multi‑head classifiers to identify astrocytes, microglia, and cancerous glioma cells.

Pipeline	Detector	Classifier	Cell‑Type Acc.
1. Single‑Stage	YOLOv8‑Seg (3 classes)	—	40 %
2. Two‑Stage CNN	YOLOv8	ResNet‑based Multi‑Head CNN	64 %
3. Two‑Stage ViT	YOLOv8	ViT‑based Multi‑Head	96 %

---

2  Dataset

The primary dataset consists of bright‑field microscopic tiles annotated by pathologists at AIIMS Jodhpur.

Class	Images
Astrocyte	181
Microglia	56
Cancerous Glioma	201

All images were resized to 512 × 512 px, normalized to ImageNet statistics, and augmented with flip, affine, and color‑jitter before training.

---

3  Methodology

3.1  Single‑Stage YOLOv8 Segmentation (Baseline)

• Goal: Direct instance segmentation & classification in one network.
• Loss: Composite (box + cls + mask).
• Limitation: Struggles to separate overlapping microglia; low F1 for minority class.

3.2  YOLOv8 + Multi‑Head CNN

1. Detection: YOLOv8 (bounding boxes only).

2. Cropping: Detected crops (128 × 128) passed to a ResNet‑50 backbone.

3. Heads:

   • Head‑A (3‑way): Astrocyte / Microglia / Cancerous.
   • Head‑B (2‑way): Cancerous / Normal.

4. Loss: CE_HA + λ · BCE_HB, λ = 0.3.

3.3  YOLOv8 + Multi‑Head Vision Transformer (Best)

• Replaces ResNet with a ViT pretrained.
• Multi‑head outputs identical to Multi-head CNN (3.2).
• Large receptive field → superior context capture → +30 pp accuracy.

---

4  Quick Start

# 1. Clone
$ git clone https://github.com/YOUR-ORG/glioma-cell-classify.git
$ cd glioma-cell-classify

# 2. Create env
$ conda create -n glioma python=3.10 && conda activate glioma
$ pip install -r requirements.txt  # ultralytics, timm, torch, streamlit …

# 3. Inference demo
$ python infer.py --weights weights/best_vit.pt --source data/sample.png

# 4. Streamlit app
$ streamlit run app.py

---

5  Repository Layout

├── app.py                # Streamlit UI
├── configs/              # YAML training configs
├── data/                 # ↳ images.zip downloaded by script
├── notebooks/            # EDA & prototyping
├── src/
│   ├── detectors.py      # YOLO wrapper
│   ├── classifiers.py    # CNN & ViT heads
│   └── train.py
└── reports/
    └── metrics.xlsx      # Full results & confusion matrices

---

6  Tech Stack

• Detection: YOLOv8
• Classification: PyTorch ▸ ResNet‑50 ▸ ViT‑B/16 (timm)
• Web Demo: Streamlit + OpenCV

---

7  How to Cite

@unpublished{sohail2024glioma,
  title     = {AI-Driven Approach for Identifying Astrocytes, Microglia, and Cancerous Glioma Cells},
  author    = {Sohail, Md. Aamir and Patel, Vishvas and Patel, Om and Maurya, Rahul},
  year      = {2024},
  note      = {Joint project, IIT Jodhpur & AIIMS Jodhpur}
}

---

8  License

This work is licensed under the MIT License – see LICENSE for details.

---

9  Acknowledgements

Special thanks to AIIMS Jodhpur Pathology Dept. for expert annotations and to HPC‑IITJ for GPU compute.

---

_{Made with ♥ at IIT Jodhpur.}