Deep Learning-Assisted Cancer Diagnosis: A CNN-Based Classification System for Lung and Colon Histopathological Images

This project was released by kknkrnwn on GitHub on June 15, 2026 (Project link: https://github.com/kknkrnwn/cancer-detection-cnn). Its core is to use convolutional neural networks (CNN) to classify lung and colon histopathological images, achieving an accuracy of 98.6% on the LC25000 dataset. It explores the application potential of AI in medical image diagnosis, aiming to provide auxiliary references for pathologists and improve diagnostic efficiency and consistency.

The original author/maintainer is kknkrnwn, and the project is published on GitHub with the title 'cancer-detection-cnn'. Cancer is one of the leading causes of death globally, and early accurate diagnosis is key to improving survival rates; traditional pathological diagnosis relies on expert experience, which is time-consuming and has subjective differences. This project focuses on the automatic classification of pathological images of two common cancers (lung and colon), explores the application of CNN in medical imaging, and assists pathologists in improving diagnostic efficiency and consistency.

The LC25000 histopathological image dataset is used, which contains scanned images of lung and colon tissue sections annotated by professional pathologists. The data can be downloaded from Kaggle. The core of the model is the CNN architecture: convolutional layers extract hierarchical features (shallow layers capture edge textures, deep layers form complex structural patterns); pooling layers reduce dimensionality and enhance translation invariance; fully connected layers combined with softmax achieve end-to-end classification.

The training model is built using TensorFlow/Keras. Data preprocessing is done via OpenCV, including size normalization and pixel value standardization; Matplotlib is used to visualize training loss curves and accuracy changes to monitor model convergence and overfitting; hyperparameters such as learning rate and batch size are tuned through experiments to balance training adequacy and generalization ability.

The model achieves a classification accuracy of 98.6% on the test set, which can reliably distinguish between cancerous and normal tissues as well as cancer types. False positive/negative samples are analyzed via confusion matrix to understand the model's strengths and weaknesses; the classification report provides precision, recall, and F1-score for each type, ensuring high recall (fewer missed diagnoses) and high precision (reducing unnecessary examinations).

The current AI system is suitable as an auxiliary tool for pathologists: quickly screening cases and marking suspicious areas to improve work efficiency; making up for talent shortages in areas with limited medical resources; serving as a quality control tool to detect abnormalities missed by humans; used in medical education to help students understand histological features; model feature learning may reveal microscopic patterns, providing new perspectives for pathological mechanism research.

Existing limitations: The LC25000 dataset does not cover all cancer types and pathological variations, and differences in quality/staining of real clinical images may affect generalization; the small sample size of some cancer types leads to weak recognition ability. Future directions: Introduce data augmentation to improve robustness; try advanced architectures such as ResNet/EfficientNet; explore attention mechanisms to enhance interpretability; conduct multi-center validation to evaluate clinical performance.

This project achieves an accuracy of 98.6% on lung and colon pathological images through the CNN classification system, verifying the feasibility of AI-assisted cancer diagnosis and laying the foundation for subsequent research and clinical applications. With technological progress and data accumulation, AI is expected to become a trusted intelligent assistant for pathologists, ultimately benefiting more patients.