# Clinical validation of lightweight CNN architectures for reliable multi-class classification of lung cancer using histopathological imaging techniques

## TL;DR

This paper compares lightweight CNN models for classifying lung cancer types from histopathological images, finding one model that achieves high accuracy with lower computational cost.

## Contribution

A reproducible framework for evaluating lightweight CNNs in lung cancer classification, identifying a model with strong generalization and reduced computational cost.

## Key findings

- Lite-V2 model achieved superior macro-F1 performance in multi-class lung cancer classification.
- The model demonstrated strong generalization on unseen test data with high accuracy.
- Lightweight CNNs can reliably classify lung cancer types with reduced computational resources.

## Abstract

Lung cancer remains one of the leading causes of cancer-related mortality worldwide, and accurate early diagnosis plays a critical role in improving patient survival. In this study, a comparative analysis of multiple lightweight Convolutional Neural Network (CNN) variants is presented for multi-class lung cancer classification using histopathological images. Four CNN architectures were designed to systematically explore the trade-off between model complexity and classification performance. Each variant was trained and evaluated within a unified experimental framework incorporating data augmentation, class balancing via computed class weights, and a custom macro-F1-based early stopping callback to ensure stable and fair performance comparison. The models were trained on three histopathological classes, Lung Benign Tissue, Lung Adenocarcinoma, and Lung Squamous Cell Carcinoma. The training process involved automated generation of accuracy, loss, and validation F1 curves, along with confusion matrices for both validation and test datasets. To assess robustness, the best-performing model was evaluated across multiple random seeds and statistical significance was established using paired McNemar’s tests against competing variants. Among the proposed variants, one model (Lite-V2) achieved superior macro-F1 performance and demonstrated strong generalization capability on unseen test data, confirming the effectiveness of lightweight CNNs in achieving high accuracy with reduced computational cost. This work highlights the potential of custom lightweight CNN architectures for efficient and reliable lung cancer classification, offering a reproducible framework that can be extended to larger datasets or adapted for clinical diagnostic applications.