wulandari, Sari Ayu Wulandari (2026) Segmentasi Citra Patch Mikroskopis Dahak Low-Light Menggunakan Hybrid Fuzzy Intuitionistic Dan Deep Learning. Doctoral thesis, Institut Teknologi Sepuluh Nopember.
![[thumbnail of 7022211023_Doctoral.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
7022211023_Doctoral.pdf - Accepted Version Restricted to Repository staff only Download (27MB) | Request a copy |
Abstract
Tuberculosis bacilli segmentation pada citra mikroskopik sputum smear merupakan tantangan penting dalam diagnosis otomatis, terutama pada kondisi low-light dan variasi pewarnaan yang menyebabkan kontras rendah serta kemiripan visual antara bacilli dan artefak. Pendekatan konvensional berbasis segmentasi piksel seringkali menghasilkan performa yang kurang stabil karena keterbatasan representasi fitur dan ketergantungan pada anotasi manual. Penelitian ini bertujuan untuk meningkatkan akurasi segmentasi bacilli melalui pengembangan metode Hybrid Residual Unet with Triplet Embedded Metric Learning (RTL) berbasis dataset patch. Dataset dibangun dari citra mikroskopik yang dipotong menjadi patch berukuran kecil, terdiri atas citra patch warna dan patch mask sebagai ground truth. Untuk meningkatkan kualitas representasi, diterapkan pendekatan multi-denoising dan strategi pseudo-annotation guna mengatasi keterbatasan anotasi manual. Selanjutnya, model RTL mengintegrasikan triplet loss pada embedding space untuk meningkatkan separabilitas fitur antara bacilli dan artefak. Eksperimen dilakukan menggunakan dataset patch mikroskopik dengan variasi kondisi iluminasi, meliputi bright, normal, dan low-light. Hasil menunjukkan bahwa metode RTL mampu meningkatkan performa segmentasi secara signifikan dengan nilai Dice sebesar 0.690 dan mIoU sebesar 0.727, serta menunjukkan robustnes terhadap variasi pencahayaan. Temuan ini menunjukkan bahwa pendekatan berbasis patch, pseudo-annotation, dan metric learning berpotensi menjadi solusi yang efektif untuk segmentasi bacilli pada citra mikroskopik dengan kondisi kompleks.
=====================================================================================================================================
Tuberculosis bacilli segmentation in sputum smear microscopy images remains a challenging task, particularly under low-light conditions and staining variability that lead to low contrast and visual similarity between bacilli and artifacts. Conventional pixel-wise segmentation approaches often exhibit unstable performance due to limited feature representation and strong dependence on manually annotated data. This study aims to improve segmentation accuracy by proposing a Hybrid Residual U-Net with Triplet Embedded Metric Learning (RTL) framework based on a patch-level dataset. The dataset is constructed by extracting small patches from microscopic images, consisting of RGB patch images and corresponding binary masks as ground truth. To address annotation scarcity and image quality issues, a multi-denoising strategy and pseudo-annotation approach are incorporated. Furthermore, the RTL model integrates triplet loss into the embedding space to enhance feature separability between bacilli and background artifacts. Experiments are conducted on patch-based microscopy datasets under varying illumination conditions, including bright, normal, and low-light scenarios. The results demonstrate that RTL achieves superior performance with a Dice score of 0.690 and mIoU of 0.727, while maintaining robustness across illumination variations. These findings indicate that the integration of patch-based representation, pseudo-annotation, and metric learning provides an effective solution for robust bacilli segmentation in complex microscopic imaging conditions.
| Item Type: | Thesis (Doctoral) |
|---|---|
| Additional Information: | - |
| Uncontrolled Keywords: | dataset patch, deep learning, low-light, pseudo-annotation, tuberculosis. |
| Subjects: | T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7878 Electronic instruments T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7882.P3 Pattern recognition systems |
| Divisions: | Faculty of Industrial Technology > Electrical Engineering > 20001-(S3) PhD Thesis |
| Depositing User: | Sari Ayu Wulandari |
| Date Deposited: | 05 Aug 2026 04:28 |
| Last Modified: | 05 Aug 2026 04:28 |
| URI: | http://repository.its.ac.id/id/eprint/143940 |
Actions (login required)
 |
View Item |