[공지사항]을 빙자한 안부와 근황 
Show more

Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

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

Created by
  • Haebom

Author

Sebastian A. Cruz Romero, Wilfredo E. Lugo Beauchamp

Outline

This paper presents a study on the development of a deep learning-based mobile application for diagnosing anemia in infants and young children in a low-resource environment. To overcome the limitations of existing anemia diagnosis methods, we present a deep learning model that detects anemia using conjunctival pallor images. We trained a MobileNet architecture-based model using the CP-AnemiC dataset (710 images), and achieved high accuracy (0.9313), precision (0.9374), and F1 score (0.9773) by applying data augmentation and cross-validation techniques. To reduce the weight of the model, we experimented with quantization of various bit widths (FP32, FP16, INT8, INT4), and confirmed that FP16 quantization was effective in reducing the model size without performance degradation. In future studies, we aim to develop a model suitable for a mobile medical environment through more efficient quantization techniques and hardware optimization.

Takeaways, Limitations

•
Takeaways:
◦
Presenting the possibility of providing an inexpensive and efficient mobile solution for anemia diagnosis in low-resource environments.
◦
Confirming the possibility of accurate and rapid anemia diagnosis through deep learning-based image analysis.
◦
Emphasizes the importance of studying quantization strategies for deploying deep learning models on mobile devices.
•
Limitations:
◦
Further validation of generalization performance is needed due to the limited dataset size (710 images).
◦
Performance degradation occurs in INT8 and INT4 quantization, requiring research on more effective quantization techniques and hardware optimization.
◦
Lack of model performance evaluation and validation in real clinical environments.
👍
Made with Slashpage