# Compact deep learning models for colon histopathology focusing performance and generalization challenges

## TL;DR

This paper introduces compact deep learning models for colon cancer histopathology classification, achieving high accuracy but facing generalization challenges on unseen data.

## Contribution

The paper introduces and evaluates four lightweight CNN models for colon histopathology classification, highlighting their performance and generalization limitations.

## Key findings

- Lite-V2 achieved near-perfect validation performance with macro-F1 ≈ 0.999 and compact size (1.53 MB).
- Lite-V2 showed significant generalization drop on the test set with macro-F1 = 0.33, indicating domain shift.
- The study emphasizes the need for domain adaptation and stain-robust training for reliable deployment.

## Abstract

Colorectal cancer is a leading cause of cancer-related mortality, and accurate analysis of histopathological images is critical for early diagnosis and improved patient outcomes. This study proposes and systematically evaluates four purpose-built lightweight convolutional neural network (CNN) variants (Lite-V0, Lite-V1, Lite-V2, and Lite-V4) for binary classification of colon histopathology images into Colon_Adenocarcinoma and Colon_Benign_Tissue. Experiments were conducted on a balanced dataset (24,000 images) with fixed train/validation/test splits and comprehensive evaluation using accuracy and macro-F1, supported by confusion matrices and ROC/precision–recall analyses. Among all variants, Lite-V2 achieved the best validation performance (macro-F1 \documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$$\approx$$\end{document} 0.999), while remaining highly compact (1.53 MB; 127,682 parameters), indicating a favorable accuracy–efficiency trade-off for deployment-oriented diagnostic support. On the independent test set, however, Lite-V2 exhibited a marked generalization drop, achieving approximately 50% accuracy and macro-F1 = 0.33, suggesting a domain-shift effect between validation and test samples. These findings demonstrate that lightweight CNNs can achieve near-perfect internal validation performance for colon histopathology classification, but robust cross-domain generalization remains essential; future work will focus on domain adaptation and stain-robust training strategies to improve reliability on unseen clinical data.