Andyartha, Putu Krisna (2025) Analisis Pemilihan Algoritma Deteksi Objek Untuk Membantu Diagnosis Peritoneal Carcinomatosis Pada Lingkungan Dengan Komputasi Terbatas. Masters thesis, Institut Teknologi Sepuluh Nopember.
![]() |
Text
6025231093-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy |
Abstract
Identifikasi letak dan ukuran tumor merupakan faktor penting dalam diagnosis kanker, termasuk pada peritoneal carcinomatosis (PC), yang menyerang selaput pelapis perut dan abdomen. Teknik konvensional menggunakan pencitraan volumetrik 3D untuk memetakan tumor pada rongga-rongga pasien, namun tergolong mahal dan membutuhkan sumber daya tinggi. Pembedahan eskploratif, khususnya laparoskopi, dapat menjadi alternatif yang lebih terjangkau, namun resolusi citra 2D cenderung lebih rendah karena batasan konteks spatial dan depth. Pendekatan berbasis deteksi objek berpotensi untuk mengatasi limitasi data 2D untuk diagnosis PC, namun pada literatur yang ada belum dioptimalkan untuk dijalankan di komputer institusi medis dengan sumber daya komputasi terbatas. Oleh karena itu, tesis ini mengajukan bantuan diagnosis PC berbasis deteksi objek untuk lingkungan dengan komputasi terbatas.Tesis ini terdiri dari tiga tahapan untuk membantu diagnosis PC. Analisis perbandingan deteksi objek dilakukan untuk memilih model deteksi objek yang efisien. Tiga belas model berbasis You Only Look Once (YOLO) dan Single-shot Detector (SSD) MobileNet dievaluasi dengan penekanan pada metrik area under precision-recall curve (AUC). Selanjutnya, model di-deploy ke dalam lingkungan dengan batasan komputasi: perangkat NVIDIA Jetson Nano dan cloud server. Model terbaik diperoleh dengan kalkulasi balance antara metrik kinerja dengan penggunaan sumber daya, yang digunakan pada tahap terakhir yakni estimasi ukuran tumor. Estimasi tumor menggabungkan data lokasi tumor dengan depth-nya yang diperoleh dari Depth Anything V2. Ukuran tumor kemudian disesuaikan berdasarkan jaraknya dengan kamera sehingga estimasi ukuran lebih akurat.Data eksperimen diperoleh dari Institut de Cancérologie de l'Ouest, pusat riset kanker di Prancis. Hasil analisis algoritma deteksi objek menunjukkan YOLOv8N unggul dengan AUC 0,936 pada batch 16 dan input size 640, dimana terhitung skor balance-nya 0,86. Model tersebut dapat berjalan konsisten dengan daya hanya 5 W. Terakhir, kombinasi data depth memungkinkan prediksi tumor terbesar pada gambar dengan akurasi 80,71%, dibandingkan 35,00% tanpa depth.
==================================================================================================================================
Identifying tumor location and sizes are crucial factors in cancer diagnosis, notably for peritoneal carcinomatosis (PC), which affects the peritoneal membrane. Conventional techniques rely on 3D volumetric scans to map tumors on patients’ cavities, however they require high costs and resources. Exploratory surgeries, especially laparoscopy, could become a more affordable alternative. However, its 2D output tends to have worse resolution due to limited spatial and depth contexts. Object detection-based approaches could overcome the limitations of 2D data, but existing literature have yet to optimize their solutions for running on medical institutions’ hardwares with limited computational resources. Therefore, this thesis aims to propose assistance for PC diagnosis with object detection optimized for resource-constrained environments.This thesis consists of three main phases to aid PC diagnosis. The first phase consist of comparative analysis to select efficient tumor detection model. Thirteen models derived from You Only Look Once (YOLO) and Single-shot Detector (SSD) MobileNet were initially evaluated with emphasis on area under precision-recall curve (AUC). Next, the models are evluated on two resource constrained environments: NVIDIA Jetson Nano and a cloud server. The best model is obtained by calculating the balance between performance and resource usage metrics, which would be used for tumor size estimation. Estimation of tumor sizes combines tumor location with depth data obtained from Depth Anything V2. Tumor sizes are adjusted based on the distance to the camera, ensuring accurate size estimation.Experimental data was provided by Institut de Cancérologie de l'Ouest, a cancer research center in France. The result of object detection algorithm analysis showed that YOLOv8N is superior with AUC of 0,936 on batch 16 and input size 640, with a calculated balance score of 0,86. The model could run consistently with only a 5 W power budget. Finally, the combination of tumor location and depth data made it possible to accurately predict the largest tumor on an image with accuracy of 80,71%, compared to just 35,00% without depth data.
Item Type: | Thesis (Masters) |
---|---|
Uncontrolled Keywords: | Deteksi Objek, Estimasi Ukuran Tumor, Komputasi Terbatas, Pemilihan Algoritma, Peritoneal Carcinomatosis, Algorithm Selection, Estimation of tumor sizes, Object Detection, Resource-constrained Environments, Peritoneal Carcinomatosis. |
Subjects: | Q Science > QA Mathematics > QA76 Computer software Q Science > QA Mathematics > QA76.585 Cloud computing. Mobile computing. Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) Q Science > QA Mathematics > QA9.58 Algorithms |
Divisions: | Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis |
Depositing User: | Putu Krisna Andyartha |
Date Deposited: | 31 Jul 2025 09:15 |
Last Modified: | 31 Jul 2025 09:15 |
URI: | http://repository.its.ac.id/id/eprint/124285 |
Actions (login required)
![]() |
View Item |