Boundary Aware Mobile U-Net dengan Multi-Stage Fine-Tuning Strategy untuk Segmentasi Rambut Pada Citra Lesi Kulit

Kostidjan, Okky Darmawan (2026) Boundary Aware Mobile U-Net dengan Multi-Stage Fine-Tuning Strategy untuk Segmentasi Rambut Pada Citra Lesi Kulit. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025232001-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (4MB) | Request a copy

Abstract

Lesi kulit adalah bagian kulit yang mengalami perubahan warna dan tekstur sehingga memiliki warna dan tekstur berbeda dari kulit di sekitarnya. Meskipun pada umumnya tidak berbahaya, lesi kulit dapat berpotensi menjadi kanker kulit. Kanker kulit merupakan salah satu jenis penyakit paling mematikan di dunia. Pesatnya perkembangan teknologi khususnya deep learning saat ini menyebabkan identifikasi dini kanker kulit lebih objektif dan cepat. Namun, metode ini memiliki kekurangan, yaitu sensitif terhadap artefak seperti rambut yang menutupi lesi kulit sehingga membutuhkan segmentasi untuk memisahkan rambut dan objek lainnya pada citra. Penelitian mengenai segmentasi rambut berbasis deep learning pada citra lesi kulit saat ini pada umumnya cenderung menggunakan metode pelatihan atau fine-tuning satu tahap dan hanya menerapkan satu loss function untuk melatik model. Pada penelitian ini, diusulkan modifikasi skip connection pada Mobile U-Net dengan melakukan integrasi Convolutional Block Attention Modul (CBAM) atau Boundary Guided Feature Fusion (BGFF) pada skip connection agar skip connection pada Mobile U-Net dapat menyaring fitur yang tepat dari encoder yang kemudian diterukan ke decoder agar bisa memprediksi mask rambut yang lebih baik. Selain itu, diusulkan juga metode pelatihan dengan menerapkan multi-stage fine-tuning strategy dan juga boundary hybrid loss function. Multi-stage fine-tuning strategy adalah metode pelatihan atau fine-tuning model yang terdiri tiga tahap dengan menggunakan loss function dan learning rate berbeda pada tiap tahapannya. Proses ini memungkinkan model mengenali variasi rambut mulai dari yang tebal, panjang, dan berwarna gelap hingga rambut yang tipis, tersebar, serta memiliki kontras rendah dan warna yang menyerupai latar belakang, baik pada area lesi kulit maupun kulit normal.
Penelitian ini bertujuan mengusulkan metode dan model segmentasi rambut pada citra lesi kulit. Tahapan penelitian ini dibagi menjadi tiga, yaitu : Pertama, menyusun loss function dan modifikasi skip connection yang sesuai untuk segmentasi rambut pada citra lesi kulit. Kedua, fine-tuning model Mobile U-Net dengan multi-stage fine-tuning strategy dan transfer learning serta ketiga, membandingkan hasil segmentasi antara metode yang diusulkan dengan metode yang sudah ada.
Evaluasi penelitian ini dilakukan pada tahap segmentasi secara kualitatif dan kuantitatif menggunakan pengukuran Dice dan Intersection over Union (IoU) dengan membandingkan kesesuaian intensitas tiap piksel citra dengan koordinat yang sama antara citra mask rambut groundtruth dengan citra mask rambut hasil prediksi. Proporsi piksel rambut yang cenderung lebih sedikit dari pada kulit sebagai latar belakangnya pada citra lesi kulit menyebabkan Dice cocok digunakan. Sebab, Dice mengukur kinerja model dengan menitik beratkan pada piksel rambut pada citra groundtruth yang diprediksi secara benar. IoU digunakan untuk mengukur kesamaan piksel pada keseluhuran citra groundtruth dan prediksi sebab metrik ini menitik beratkan pengukuran kinerja model pada akurasi prediksi model pada tepian objek.
Hasil dari penelitian ini menunjukkan bahwa gabungan multi-stage fine-tuning dapat meningkatkan kinerja Mobile U-Net dari nilai Dice 76.59% dan IoU 62.11% meningkat menjadi 77.84% (Dice) dan 63.74% (IoU). Selain itu, integrasi BGFF meningkatkan kinerja Mobile U-Net yang sudah dilatih dengan multi-stage fine-tuning dan boundary hybrid loss sehingga dapat mencapai nilai Dice 79.08% dan IoU sebesar 65.42%
==============================================================================================================================================
Skin lesions are skin parts that experience color and texture changes so that they are different from the surrounding skin. Although generally harmless, skin lesions can potentially become skin cancer. Skin cancer is one of the most deadly diseases in the world. The swift technology development, especially deep learning, currently makes early skin cancer identification more objective and faster. However, this method has disadvantages, namely being sensitive to artifacts such as hair that covers skin lesions so that it requires segmentation to separate hair and other objects. Existing studies on deep learning based hair segmentation in skin lesion images predominantly employ single-stage training or fine-tuning strategies and rely on a single loss function during model training. In this study, an enhanced skip-connection mechanism for Mobile U-Net is proposed by integrating either a Convolutional Block Attention Module (CBAM) or a Boundary Guided Feature Fusion (BGFF) module into the skip connections. This integration enables the skip connections to selectively filter relevant features from the encoder before transmitting them to the decoder, thereby improving the reconstruction quality of hair masks. In addition, a training method is also proposed by applying a multi-stage fine-tuning strategy and boundary aware hybrid loss function. The multi-stage fine-tuning strategy is a three-stage training or fine-tuning method, each using a different loss function and learning rate. This process allows the model to recognize hair variations ranging from thick, long, and dark to thin, scattered hair with low contrast and a color similar to the background, both in skin lesions and normal skin.
This research aims to propose an effective method and model for hair segmentation in skin lesion images. The research workflow is divided into three main stages: first, designing an appropriate combination of loss functions and modifying skip connections for hair segmentation in skin lesion images; second, fine-tuning the Mobile U-Net model using a multi-stage fine-tuning strategy and transfer learning; and third, conducting hair segmentation performance comparisons between the proposed method and existing approaches.
The evaluation of this study is conducted at the segmentation stage through both qualitative and quantitative assessments using the Dice Similarity Coefficient (Dice) and Intersection over Union (IoU). These metrics are computed by comparing the suitability of the the pixel intensities at corresponding spatial coordinates between the ground-truth hair segmentation masks and the predicted segmentation results. In skin lesion images, the proportion of hair pixels is considerably smaller than that of the background skin, making Dice an appropriate metric for this task. This is because Dice emphasizes the correct prediction of hair pixels in the ground-truth masks and is less influenced by the dominant background regions. Meanwhile, IoU is employed to measure the overall pixel-wise similarity between the ground-truth and predicted segmentation masks, as this metric places greater emphasis on the accuracy of object boundary delineation, thereby providing a stricter evaluation of segmentation performance at the edges of hair structures.
The experimental results demonstrate that the combination of multi-stage fine-tuning improves the performance of Mobile U-Net, increasing the Dice score from 76.59% and the IoU from 62.11% to 77.84% (Dice) and 63.74% (IoU). Furthermore, the integration of BGFF further enhances the performance of Mobile U-Net trained with multi-stage fine-tuning and boundary hybrid loss, achieving a Dice score of 79.08% and an IoU of 65.42%.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Boundary aware hybrid loss, Mobile U-Net, multi-stage fine-tuning, lesi kulit, segmentasi, segmentation, skin lesion
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Okky Darmawan Kostidjan
Date Deposited:	26 Jan 2026 03:56
Last Modified:	26 Jan 2026 03:56
URI:	http://repository.its.ac.id/id/eprint/130307

Actions (login required)

View Item