Klasifikasi Tumor Payudara Menggunakan Arsitektur Dual-Stream Berbasis Segmentasi Hibrida YOLO dan MedSAM2 pada Citra Ultrasonografi

Azhar, Daffa Muhamad (2026) Klasifikasi Tumor Payudara Menggunakan Arsitektur Dual-Stream Berbasis Segmentasi Hibrida YOLO dan MedSAM2 pada Citra Ultrasonografi. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025241057-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (4MB) | Request a copy

Abstract

Deteksi dini kanker payudara melalui ultrasonografi (USG) sangat krusial, namun rentan terhadap subjektivitas akibat derau (speckle noise). Sistem Computer-Aided Diagnosis (CAD) menjanjikan solusi objektif, tetapi pendekatan Multi-Task Learning (MTL) saat ini kerap mengorbankan kualitas segmentasi demi akurasi klasifikasi. Selain itu, implementasi Foundation Model masih terhambat oleh kebutuhan panduan (prompt) manual dan risiko kegagalan sistem (single point of failure) jika pelokalan awal meleset. Penelitian ini mengusulkan kerangka kerja diagnostik hibrida sekuensial otomatis. Pada tahap pelokalan, arsitektur pendeteksi dioptimasi menggunakan fitness function kustom berorientasi Recall (YOLOv9) untuk menekan angka false negative secara agresif. Kotak pembatas dari detektor ini kemudian memandu MedSAM2 secara otomatis pada tahap segmentasi. Untuk mencegah kegagalan deteksi pada lesi yang tersamar, diterapkan mekanisme fallback proksi berbasis titik sentroid statistik guna menjaga keberlanjutan sistem. Pada tahap klasifikasi, arsitektur Dual-Stream berbasis ResNet50 mengekstraksi fitur global dan lokal dari masker MedSAM2. Fitur ini diintegrasikan melalui Atrous Spatial Pyramid Pooling (ASPP) dan disaring oleh Convolutional Block Attention Module (CBAM). Optimasi model akhir menggunakan fungsi validation fitness kustom (AUC dan F2-Score) untuk memprioritaskan sensitivitas klinis. Evaluasi menggunakan 5-fold cross-validation pada himpunan data BUSI menunjukkan keunggulan sistem usulan. Pada segmentasi, kerangka kerja ini mencapai rata-rata Dice 0,8841, IoU 0,8148, dan Recall 0,8976. Pada klasifikasi, jaringan Dual-Stream mencetak Area Under Curve (AUC) 0,9602, Akurasi 0,8980, dan sensitivitas tinggi sebesar 0,9286, melampaui performa berbagai metode state-of-the-art (SOTA) dan pendekatan MTL. Kesimpulannya, arsitektur hibrida ini sukses mengatasi kompromi fungsional MTL dan kelemahan pelokalan tunggal. Penelitian mendatang disarankan untuk memvalidasi sistem pada himpunan data eksternal (multisentris), meningkatkan kinerja arsitektur detektor (YOLO) sebagai penghasil prompt otomatis, serta mengeksplorasi fungsi kerugian (loss function) alternatif guna mencapai batas atas (upper bound) performa model secara utuh.
=======================================================================================================================================
Early detection of breast cancer through ultrasonography (US) is crucial, yet highly susceptible to subjectivity caused by speckle noise. Computer-Aided Diagnosis (CAD) systems offer a promising objective solution, but current Multi-Task Learning (MTL) approaches often compromise segmentation quality in favor of classification accuracy. Furthermore, the implementation of Foundation Models is still hindered by the need for manual prompts and the risk of a single point of failure if the initial localization misses the target. This study proposes an automated sequential hybrid diagnostic framework. In the localization phase, the detection architecture is optimized using a custom Recall-oriented fitness function (YOLOv9) to aggressively suppress the false negative rate. Bounding boxes from this detector then automatically guide MedSAM2 during the segmentation phase. To prevent detection failure on obscured lesions, a statistical centroid-based proxy fallback mechanism is implemented to maintain system continuity. In the classification phase, a ResNet50-based Dual-Stream architecture extracts global and local features from the MedSAM2 masks. These features are integrated through Atrous Spatial Pyramid Pooling (ASPP) and refined by a Convolutional Block Attention Module (CBAM). The final model optimization utilizes a custom validation fitness function (AUC and F2-Score) to prioritize clinical sensitivity. Evaluation using 5-fold cross-validation on the BUSI dataset demonstrates the superiority of the proposed system. In segmentation, the framework achieves an average Dice of 0.8841, IoU of 0.8148, and Recall of 0.8976. In classification, the Dual-Stream network scores an Area Under Curve (AUC) of 0.9602, an Accuracy of 0.8980, and a high sensitivity of 0.9286, outperforming various state-of-the-art (SOTA) methods and MTL approaches. In conclusion, this hybrid architecture successfully overcomes the functional compromises of MTL and the weaknesses of standalone localization. Future research should validate the system on external (multicentric) datasets, enhance the performance of the detector architecture (YOLO) as an automated prompt generator, and explore alternative loss functions to fully reach the model's performance upper bound.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Klasifikasi Dual-Stream, MedSAM2, Segmentasi Citra, Tumor Payudara, YOLOv9, Breast Tumor, Dual-Stream Classification, Image Segmentation
Subjects:	Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines. Q Science > Q Science (General) > Q337.5 Pattern recognition systems Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Daffa Muhamad Azhar
Date Deposited:	27 Jul 2026 07:37
Last Modified:	27 Jul 2026 07:37
URI:	http://repository.its.ac.id/id/eprint/138022

Actions (login required)

View Item