Lung Cancer Histopathology Classification with Basic CNN & DenseNet-121

Introduction

Lung cancer is one of the leading causes of cancer deaths, and catching it early matters enormously. A key step is examining tissue samples under a microscope, but doing this by eye is slow, tiring, and depends on scarce specialist time.

Image software can support this work by learning the tell-tale patterns in tissue images, helping screen samples faster and label them more consistently. This project builds a system that sorts lung tissue images into healthy tissue and two common cancer types.

Problem Statement

Reading tissue slides reliably is challenging:

1. Specialists must spot subtle differences across thousands of slides.
2. Results can vary between readers: and early cancer can resemble harmless growths.
3. The two cancer types look different from each other and need to be told apart.
4. Any helper needs to run on modest hardware without losing reliability.

Objectives

1. Sort lung tissue images into healthy tissue and two cancer types.
2. Compare a few image models to see which is most reliable.
3. Use a balanced: well-prepared image set for fair training.
4. Judge the models on clear: side-by-side measures.
5. Find the approach best suited to supporting early detection.

Dataset

The work is built around a balanced set of lung tissue images covering healthy tissue and two cancer types. Depending on the client, this can be a public medical-imaging collection, licensed or private pathology data, images from other trusted sources, or a custom set gathered for the project. Images are cleaned, balanced, and lightly expanded with standard adjustments before training, with privacy respected throughout.

Methodology

The project trains and compares four image models on the same prepared data: Basic CNN, Modified AlexNet, VGG19, and DenseNet-121, from a simple built-from-scratch network to larger, more established designs, each learning the visual patterns that separate healthy tissue from the two cancer types.

The value here is practicality alongside accuracy. Rather than assuming the biggest model wins, it weighs reliability against how much computing power each design needs, pointing toward an approach that is dependable enough to assist specialists yet light enough to run in modest lab settings.