Performance Analysis of Post-Training Quantization for CNN-based Conjunctival Pallor Anemia Detection

TL;DR

This study evaluates the impact of post-training quantization on MobileNet-based CNN models for anemia detection from conjunctival images, balancing model size and accuracy for mobile health deployment.

Contribution

It provides a comprehensive analysis of quantization effects on CNN performance for anemia detection, highlighting optimal schemes for edge device deployment.

Findings

FP16 quantization maintains high accuracy and F1 score.

INT8 and INT4 quantization significantly reduce model performance.

Post-training quantization enables efficient mobile health applications.

Abstract

Anemia is a widespread global health issue, particularly among young children in low-resource settings. Traditional methods for anemia detection often require expensive equipment and expert knowledge, creating barriers to early and accurate diagnosis. To address these challenges, we explore the use of deep learning models for detecting anemia through conjunctival pallor, focusing on the CP-AnemiC dataset, which includes 710 images from children aged 6-59 months. The dataset is annotated with hemoglobin levels, gender, age and other demographic data, enabling the development of machine learning models for accurate anemia detection. We use the MobileNet architecture as a backbone, known for its efficiency in mobile and embedded vision applications, and fine-tune our model end-to-end using data augmentation techniques and a cross-validation strategy. Our model implementation achieved an accuracy of 0.9313, a precision of 0.9374, and an F1 score of 0.9773 demonstrating strong performance on the dataset. To optimize the model for deployment on edge devices, we performed post-training quantization, evaluating the impact of different bit-widths (FP32, FP16, INT8, and INT4) on model performance. Preliminary results suggest that while FP16 quantization maintains high accuracy (0.9250), precision (0.9370), and F1 Score (0.9377), more aggressive quantization (INT8 and INT4) leads to significant performance degradation. Overall, our study supports further exploration of quantization schemes and hardware optimizations to assess trade-offs between model size, inference time, and diagnostic accuracy in mobile healthcare applications.