Deteksi Dini Kanker Serviks Menggunakan Model Berbasis CNN dengan Quantization dan Explainable AI

Luqmanulhakim, Naufan Zaki (2026) Deteksi Dini Kanker Serviks Menggunakan Model Berbasis CNN dengan Quantization dan Explainable AI. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5027221065-Undergraduate_Thesis.pdf Download (6MB)

Abstract

Kanker serviks merupakan salah satu penyebab utama kematian pada perempuan, khususnya di wilayah dengan keterbatasan akses layanan kesehatan, sementara metode skrining Inspeksi Visual Asam Asetat (IVA) masih bergantung pada subjektivitas tenaga medis dalam proses interpretasi. Penelitian ini bertujuan mengembangkan sistem deteksi dini kanker serviks berbasis citra tes IVA menggunakan pendekatan Convolutional Neural Network (CNN) yang diimplementasikan secara offline melalui mekanisme on-device inference pada perangkat Android. Tiga arsitektur CNN berbasis transfer learning, yaitu ResNet-50, EfficientNet-B0, dan MobileNetV3-Small dievaluasi secara komparatif melintasi sembilan skenario ablation study strategi preprocessing. Guna menekan bahaya klinis shortcut learning akibat efek Clever Hans dari elemen non-biologis latar belakang, diterapkan Teknik pemotongan otomatis berbasis kontur (OpenCV Auto-Crop). Model dioptimasi menggunakan Post-Training Quantization (PTQ) ke format TensorFlow Lite INT8 untuk meningkatkan efisiensi komputasi pada perangkat bergerak. Hasil eksperimen menunjukkan bahwa kombinasi OpenCV Auto-Crop (skenario NB-6) dengan arsitektur ResNet-50 INT8 merupakan model final terbaik, berhasil mencapai keseimbangan performa optimal dengan nilai PR-AUC sebesar 0,759, sensitivitas (recall) 89%, spesifisitas 56%, serta akurasi keseluruhan sebesar 75,7% pada hold-out test set. Model ini secara signifikan mengungguli model baseline machine learning tradisional (handcrafted features + Naive Bayes) yang hanya meraih akurasi total 44,8%. Optimasi kuantisasi berhasil memangkas ukuran model menjadi 23,12 MB dengan rata-rata waktu inferensi lokal yang sangat responsif, yaitu 110,84 ms pada perangkat Android. Model terbaik ini diintegrasikan dengan metode Explainable AI (XAI) berbasis Grad-CAM untuk memvisualisasikan area patologis (acetowhite) secara transparan guna memvalidasi keputusan klinis. Hasil pengujian User Acceptance Testing (UAT) dan validasi klinis lapangan bersama 4 dokter spesialis dan residen Obstetri dan Ginekologi di RSUP Hasan Sadikin Bandung mengonfirmasi kelayakan prototipe aplikasi CervicalAI sebagai alat bantu skrining lapis pertama yang portabel dan independen dari konektivitas internet.
================================================================================================================================
Cervical cancer remains one of the leading causes of cancer-related mortality among women, particularly in regions with limited access to healthcare services, while Visual Inspection with Acetic Acid (VIA) screening still relies heavily on subjective interpretation by medical personnel. This research aims to develop an early detection system for cervical cancer based on VIA images using a Convolutional Neural Network (CNN) implemented through offline on-device inference on Android devices. Three transfer learning-based CNN architectures, namely ResNet-50, EfficientNet-B0, and MobileNetV3-Small, were comparatively evaluated across nine ablation study scenarios of preprocessing strategies. To mitigate the clinical risks of shortcut learning induced by the Clever Hans effect from non-biological background artifacts, an automated contour-based cropping technique (OpenCV Auto-Crop) was implemented. The models were optimized using Post-Training Quantization (PTQ) into TensorFlow Lite INT8 format to enhance computational efficiency on mobile devices. Experimental results demonstrate that the combination of OpenCV Auto-Crop (scenario NB-6) and the ResNet-50 INT8 architecture emerged as the optimal final model, achieving a balanced peak performance with a PR-AUC of 0.759, sensitivity (recall) of 89%, specificity of 56%, and an overall accuracy of 75.7% on the hold-out test set. This model significantly outperformed the traditional machine learning baseline (handcrafted features + Naive Bayes), which only achieved an overall accuracy of 44.8%. The quantization optimization successfully compressed the model size to 23.12 MB while delivering a highly responsive local inference latency of 110.84 ms on the Android device. The selected model was further integrated with a Grad-CAM-based Explainable AI (XAI) method to transparently visualize pathological (acetowhite) regions for clinical decision validation. User Acceptance Testing (UAT) and field clinical validation conducted with 4 Obstetrics and Gynecology specialists and residents at RSUP Hasan Sadikin Bandung confirmed the feasibility of the CervicalAI prototype as a portable, offline-capable first-line screening decision support tool.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Convolutional Neural Network, Explainable AI, Grad-CAM, Kanker Serviks, OpenCV Auto-Crop, Post-Training Quantization, Tes IVA, Cervical Cancer, Convolutional Neural Network, Explainable AI, Grad-CAM, OpenCV Auto-Crop, Post-Training Quantization, VIA Test
Subjects:	T Technology > T Technology (General) > T57.5 Data Processing T Technology > T Technology (General) > T57.8 Nonlinear programming. Support vector machine. Wavelets. Hidden Markov models.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Information Technology > 59201-(S1) Undergraduate Thesis
Depositing User:	Naufan Zaki Luqmanulhakim
Date Deposited:	15 Jul 2026 01:55
Last Modified:	15 Jul 2026 01:55
URI:	http://repository.its.ac.id/id/eprint/134959

Actions (login required)

View Item