Segmentasi Lesi Melanositik Dengan Menerapkan Semi-supervised Generative Adversarial Network

Sinaga, Ariel Duftin (2025) Segmentasi Lesi Melanositik Dengan Menerapkan Semi-supervised Generative Adversarial Network. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5002211046-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (12MB)

Abstract

Segmentasi pada citra medis, khususnya citra lesi melanositik, merupakan langkah krusial dalam mendeteksi kanker kulit secara dini. Salah satu pendekatan yang umum digunakan dalam melakukan segmentasi pada citra lesi melanositik adalah dengan metode secara fully supervised dengan menggunakan beberapa model segmentasi yang berbasis jaringan konvolusi. Namun, ketersediaan ground truth sering kali terbatas dan tidak sebanding dengan jumlah citra lesi yang tersedia, sehingga menjadi kendala dalam proses pelatihan model. Untuk mengatasi permasalahan tersebut, penelitian Tugas Akhir ini menggunakan metode semi-supervised berbasis Generative Adversarial Network (GAN) yang mampu meningkatkan kualitas segmentasi meskipun hanya menggunakan sebagian kecil data berlabel atau yang memiliki ground truth dan tetap memanfaatkan citra tanpa label. Hasil pengujian menunjukkan bahwa metode semi-supervised berbasis Generative Adversarial Network mampu memberikan hasil segmentasi yang lebih baik dibandingkan pendekatan fully supervised yang hanya mengandalkan data berlabel terbatas dengan nilai Average Hausdorff Distance terbaik sebesar 30.1524, nilai Dice Similarity Coefficient atau Dice Score terbaik sebesar 0.8062 dan nilai IoU (Intersection over Union) terbaik sebesar 0.6753.
================================================================================================================================
Segmentation of medical images, particularly images of melanocytic lesions, is a crucial step in the early detection of skin cancer. One common approach used in segmenting images of melanocytic lesions is the fully supervised method, which uses several segmentation models based on convolutional networks. However, the availability of ground truth is often limited and does not match the number of lesion images available, which poses a challenge in the model training process. To address this issue, this thesis research uses a semi-supervised method based on Generative Adversarial Network (GAN) that can improve segmentation quality even when using only a small portion of labeled data or data with ground truth, while still utilizing unlabeled images. Test results show that the semi-supervised method based on Generative Adversarial Network is able to provide better segmentation results than the fully supervised approach, which only relies on limited labeled data, with the best Average Hausdorff Distance value of 30.1524, the best Dice Similarity Coefficient or Dice Score value of 0.8062, and the best IoU (Intersection over Union) value of 0.6753.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Segmentasi Citra, Fully Supervised, Semi-supervised, Generative Adversarial Network, Image Segmentation, Fully Supervised, Semi-supervised, Generative Adversarial Network
Subjects:	Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Mathematics > 44201-(S1) Undergraduate Thesis
Depositing User:	Ariel Duftin Sinaga
Date Deposited:	01 Aug 2025 01:52
Last Modified:	01 Aug 2025 01:52
URI:	http://repository.its.ac.id/id/eprint/125327

Actions (login required)

View Item