Project Motivation

Lung cancer is one of the deadliest cancers worldwide, and timely, accurate diagnosis is critical. This project explores the use of artificial intelligence in histopathological image classification to support clinical decision-making and improve outcomes. By leveraging deep learning, we aim to assist pathologists in detecting lung cancer subtypes more consistently and efficiently.

The project began with a convolutional neural network model trained to detect lung nodules from chest X-ray images. However, to improve resolution, detail, and diagnostic relevance, the focus pivoted toward histopathology. This led to the integration of LC25000 and LungHist7000 datasets for model training on real stained lung tissue slides.

Project Goals

This web-based diagnostic prototype was developed with the following objectives:

  • Train an EfficientNet-B0-based CNN capable of identifying benign lung tissue, adenocarcinoma, and squamous cell carcinoma.
  • Augment limited datasets using controlled transformations to prevent overfitting.
  • Deliver real-time image classification via a clean, interactive web interface.
  • Enhance model explainability with integrated saliency maps and Grad-CAM overlays.
  • Lay a foundation for scalable clinical tools that prioritize both performance and transparency.

Development Journey

Initial Concept

Started as a chest X-ray lung cancer detection model using a small dataset of annotated radiographs.

Pivot to Histopathology

Recognized the clinical limitations of X-ray and transitioned to histological slide classification using LC25000.

Dataset Expansion

Incorporated real clinical slides from LungHist7000 to strengthen model generalization and real-world relevance.

Model Engineering

Used EfficientNet-B0 as the backbone. Applied preprocessing, data augmentation, and fine-tuning for optimal accuracy.

Web App Integration

Developed a user-friendly Flask interface with diagnostic feedback, saliency visualization, and Grad-CAM support.

Author & Links

This project was developed by Liviu Orehovschi as part of an academic exploration in machine learning and medical image processing. It combines research, development, and user-centered design into a functional demo of how AI can enhance cancer diagnostics.

GitHub: github.com/liviuorehovschi
LinkedIn: linkedin.com/in/liviuorehovschi