LUCIAN — Lung Carcinoma Histopathology Imaging & Analysis

Autonomous Tissue Classification, Grad-CAM Explainability, and Intelligence-Driven Cancer Diagnostics.

---

Overview

Lung cancer remains a leading cause of cancer-related mortality globally. Early and precise classification of tissue types — Adenocarcinoma (LUAD), Squamous Cell Carcinoma (LUSC), and Benign tissue — is fundamental for effective treatment planning.

LUCIAN is a specialized Computer Vision system that automates the analysis of histopathology slides. By leveraging SOTA deep learning architectures and explainability frameworks, this project bridges the gap between raw medical imaging and actionable diagnostic insights.

Technical Features

• Advanced Deep Learning: Powered by a fine-tuned ConvNeXt-Base model, achieving high accuracy on standardized histopathology datasets.
• Visual Explainability: Integrated Gradient-weighted Class Activation Mapping (Grad-CAM) to visualize high-impact regions in tissue slides.
• Systematic Methodology: Developed following the CRISP-DM framework, ensuring a data-driven approach from business understanding to deployment.
• Production-Ready CI/CD: Automated testing (pytest) and deployment pipelines ensuring consistent application updates and reliability.
• Flexible Deployment: Multi-platform support including Docker, Streamlit Cloud, Hugging Face Spaces, and Azure.

Technology Stack

Core Intelligence

• Deep Learning: TensorFlow 2.19, Keras
• Processing: NumPy, Pandas
• Explainability: Grad-CAM (OpenCV + TF)

Web Interface

• Framework: Streamlit 1.45 (Multi-page Architecture)
• Styling: Vanilla CSS integration
• Metrics Visualization: Plotly

Infrastructure & DevOps

• Containerization: Docker
• Version Control & CI/CD: Git, GitHub Actions
• Model Registry: Hugging Face Hub

Model Architecture

The core of LUCIAN is a transfer learning implementation using the ConvNeXt-Base backbone, pre-trained on ImageNet-1K. The model is optimized for 224 x 224 px histopathology inputs.

• Base Model: ConvNeXt-Base (feature extractor)
• Classification Head:
  • Global Average Pooling
  • Dense Layer (256 units, ReLU activation)
  • Dropout (30% rate for regularization)
  • Output Layer (3 units, Softmax activation)
• Dataset: Subset of LC25000 Lung Histopathology Dataset (3,000 images across 3 classes)

System Architecture

graph TD
    subgraph Data_Input [Ingestion Level]
        Image[Histopathology Slide] -->|Upload| Web[Streamlit UI]
    end
    
    subgraph Intelligence_Level [Intelligence Level]
        Web -->|Input Tensors| Model[ConvNeXt-Base Engine]
        Model -->|Gradient Flow| GradCAM[Grad-CAM Visualizer]
        Model -->|Softmax Out| Result[Class Confidence]
    end
    
    subgraph Deployment_Level [Gateway & Delivery]
        UI[Streamlit Dashboard] -->|User Interaction| API[Internal Logic]
        API -->|Fetch Model| HF[Hugging Face Hub]
        CI[GitHub Actions] -->|Sync Deployment| HFS[HF Spaces / Docker]
    end

Loading

---

Performance & Metrics

LUCIAN was evaluated using multiple data split strategies to ensure robustness and generalization.

Core Metrics (80:10:10 Split)

Parameter	Value	Description
Test Accuracy	93.67%	Final performance on unseen test data
F1-Score (Macro)	93.64%	Balanced precision and recall indicator
Precision (Macro)	93.63%	Accuracy of positive predictions
Recall (Macro)	93.67%	Ability to identify all relevant instances

Per-Class Performance

Class	Precision	Recall	F1-Score
Adenocarcinoma	91.75%	89.00%	90.36%
Benign Tissue	98.04%	100.00%	99.01%
Squamous Carcinoma	91.09%	92.00%	91.54%

---

Deployment Guide

Prerequisites

• Python 3.12+
• Docker (optional for containerized runs)

Execution Options

Option 1: Local Development Mode

# Initialize environment
python -m venv venv
source venv/bin/activate  # venv\Scripts\activate on Windows
pip install -r requirements.txt

# Run Application
streamlit run app.py

Option 2: Docker Deployment

# Build Container
docker build -t lucian .

# Run Container
docker run -p 7860:7860 lucian

Configuration

The application centralizes its constants and model settings within src/config.py. Key configurable parameters include:

• HF_REPO_ID: The Hugging Face Hub repository containing the model weights.
• IMAGE_SIZE: Default tensor dimensions (Current: 224x224).
• GRADCAM_LAYER: Target layer for explainability heatmaps (Current: flatten).

---

Author

Felix Hardyan

• HuggingFace: felixhrdyn

---

Undergraduate Thesis Project — Computer Science, 2025