# Lightweight Deep Learning Model for Classification of Normal and Abnormal Vasculature in Organoid Images

## TL;DR

This paper introduces a fast and accurate deep learning model to automatically classify normal and abnormal blood vessel patterns in organoid images, improving efficiency and consistency in research.

## Contribution

A lightweight deep learning model for real-time vasculature classification in organoids with high accuracy and improved inference speed.

## Key findings

- Modified EfficientNet models achieved 0.90, 0.99, and 1.00 accuracy for vasculature classification.
- The models processed images at 51.1 to 32.4 FPS on CPU, showing 70% faster inference than original models.
- Data augmentation and noise addition helped address class imbalance in the dataset.

## Abstract

Human organoids are 3D cell culture models that precisely replicate the microenvironment of real organs. In organoid-based experiments, assessing whether the internal vasculature has formed normally is essential for ensuring the reliability of experimental results. However, conventional vasculature assessment relies on manual inspection by researchers, which is time-consuming and prone to variability caused by subjective judgment. This study proposes a lightweight deep learning model for automatic classification of normal and abnormal vasculature in vascular organoid images. The proposed model is based on EfficientNet by replacing the activation function SiLU with ReLU and removing the Squeeze-and-Excitation (SE) blocks to reduce computational complexity. The dataset consisted of vascular organoid images obtained from co-culture experiments. Data augmentation and noise addition were performed to alleviate class imbalance. Experimental results show that the proposed Modified 3 models (B0, B1, B2) achieved accuracy of 0.90, 0.99, and 1.00, respectively, with corresponding inference speed of 51.1, 36.0, and 32.4 FPS on the CPU, demonstrating real-time inference capability and an average speed improvement of 70% compared to the original models. This study presents an efficient automated analysis framework that enables quantitative and reproducible vasculature assessment by introducing a lightweight model that maintains high accuracy and supports real-time processing.