Abstract

This research presents an automated approach to classifying C. elegans roundworms using convolutional neural networks. Using a fully convolutional U-net with a ResNet encoder, we performed semantic image segmentation on microscopy images from the Broad Bioimage Benchmark Collection. By labeling pixels as “alive,” “dead,” or “background,” we accomplished two key classification tasks: determining whether entire images contain roundworms treated with amoxicillin, and classifying individual roundworms as dead or alive. Results demonstrate 100% accuracy in image-level classification and 93% accuracy in individual organism classification.

Introduction

Caenorhabditis elegans is a valuable model organism for testing anti-infection drug therapies. When studying ampicillin effectiveness against bacterial infections, treated roundworms exhibit a curved, smooth “alive” phenotype, while untreated ones show a straight, uneven “dead” phenotype.

This research automates two classification tasks: determining if images contain treated C. elegans, and classifying individual roundworms as alive or dead based on morphological appearance using semantic segmentation with a U-net architecture.

Methodology

Dataset: 100 grayscale microscopy images from the Broad Bioimage Benchmark Collection (52 untreated, 48 treated), containing 10-20 roundworms each. Images were resized from 696×520 to 704×512 pixels for model compatibility.

Data Preparation: Images split 80/20 for training/testing. Over 1,500 individual C. elegans were manually labeled using curvature heuristics, with pixels classified as background (0), alive (1), or dead (2).

Model: U-net with ResNet encoder trained for 10 epochs on Stanford’s Sherlock computing cluster. The semantic segmentation model outputs pixel-wise classifications for entire images.

Evaluation: Image-level classification based on pixel count ratios (more “alive” pixels = untreated, more “dead” pixels = treated). Individual organism classification determined by dominant pixel values in each roundworm region.

Results

Image-Level Classification: Achieved 100% accuracy, correctly identifying all 20 test images as treated or untreated.

Individual Organism Classification: Achieved 93% accuracy (255/275 organisms correctly classified), with 94% sensitivity for alive classification and 92.4% sensitivity for dead classification.

Discussion

The results provide strong preliminary evidence for automated C. elegans classification using deep learning. The perfect image-level classification and high individual organism accuracy demonstrate the viability of this approach for high-throughput drug screening applications.

Limitations: The small dataset (100 images) limits robustness assessment, and some hyperparameters require optimization. Manual labeling introduces potential subjective bias.

Future Work: Plans include parameter optimization, dataset expansion through GANs and data augmentation, and comparative analysis with existing methods.

Significance

This work demonstrates that semantic segmentation can effectively automate C. elegans classification, offering potential for scalable drug screening applications while providing detailed pixel-level analysis of organism morphology.

Research conducted or for a at Stanford CS 279 Project, with computational support from the Stanford Sherlock computing cluster. Github Link, contains some the project notebook with writeup, code, and figures.