On-Device Deployment of Cerviray AI: Optimization via Knowledge Distillation and Quantization for Mobile Clinical Environments

Byeongin Moon, Jaeyun Song, Dongha Lee, Seongmin Kim, Donghoon Suh, Hansol Choi, JAEHOON JEONG

Keywords: Cervical Cancer, On-device AI, Knowledge Distillation, Model Quantization, Vision Transformer

TL;DR: We present an edge-optimized cervical cancer screening system, Cerviray AI, achieving 97.61% accuracy and 3.4 s inference on a tablet CPU via knowledge distillation and INT8 quantization, enabling accessible screening in resource-limited settings.

Registration Requirement: Yes

Abstract: Artificial intelligence has significantly advanced the diagnostic accuracy of Visual Inspection with Acetic acid (VIA) for cervical cancer screening. However, to overcome the GPU dependency of deep learning models and ensure their applicability in point-of-care settings, we present an on-device version of Cerviray AI. By employing Knowledge Distillation (ViT-Base to ViT-Tiny) and INT8 Post-Training Quantization, we successfully migrated the system from an RTX 4060 GPU to a Samsung Galaxy Tab S7 CPU. The optimized model achieves a clinical-grade accuracy of 97.61\% (only a 0.24\% drop) and an inference speed of 3.4s per image. This work demonstrates the potential of edge-AI in democratizing high-fidelity cancer screening for resource-limited, decentralized settings.

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 25