Pengembangan Model Deteksi Subtipe Acute Lymphoblastic Leukemia pada Dataset Citra Mikroskopis yang Jumlahnya Terbatas, Bervariasi dan Tumpang Tindih

Mustaqim, Tanzilal (2025) Pengembangan Model Deteksi Subtipe Acute Lymphoblastic Leukemia pada Dataset Citra Mikroskopis yang Jumlahnya Terbatas, Bervariasi dan Tumpang Tindih. Doctoral thesis, Institut Teknologi Sepuluh Nopember.

Text 7025222005-Desertasion.pdf Restricted to Repository staff only Download (10MB) | Request a copy

Abstract

Analisis citra mikroskopis sel darah mempunyai peran penting dalam diagnosis subtipe Acute Lymphoblastic Leukemia (ALL). Metode-metode sebelumnya, seperti Convolutional Neural networks (CNN) dan teknik pengolahan citra tradisional, telah digunakan untuk deteksi objek pada citra mikroskopis. Namun, metode-metode tersebut masih menghadapi keterbatasan dalam integrasi fitur lokal dan global serta belum optimal dalam menangani variasi visual citra, seperti sel yang saling tumpang tindih dan ketidakseimbangan dataset. Oleh karena itu, penelitian ini bertujuan untuk mengembangkan metode deteksi objek yang lebih akurat pada citra mikroskopis untuk deteksi subtipe ALL dengan mengatasi beberapa gap penelitian, termasuk keterbatasan dan ketidakseimbangan dataset serta integrasi fitur map lokal dan global. Penelitian ini dibagi menjadi beberapa tahap utama. Tahap pertama fokus pada pengembangan metode data augmentasi berbasis superpixel untuk mengatasi keterbatasan dan ketidakseimbangan dalam dataset citra mikroskopis sel darah subtipe ALL. Metode usulan ini memanfaatkan teknik seperti Gradient-weighted Class Activation Mapping (Grad-CAM), Multi-Head Attention, SaliencyMix, dan OcCaMix untuk meningkatkan keragaman dan representativitas data tanpa menambah jumlah data training yang ada. Hasilnya, metode FocusAugMixV1 mencapai akurasi sebesar 99,07%, jauh melampaui metode sebelumnya yang berkisar antara 65% hingga 97,52%. Tahap kedua melibatkan pengembangan metode image enhancement adaptive menggunakan Contrast Limited Adaptive Histogram Equalization (CLAHE) untuk meningkatkan kualitas citra mikroskopis dengan menyesuaikan kontras dan pewarnaan secara adaptif, sehingga mengurangi artefak dan kompleksitas proses. Tahap ketiga penelitian ini berfokus pada optimasi hyperparameter model deteksi objek YOLOv8 menggunakan Grey Wolf Optimizer (GWO) dan modifikasi arsitektur YOLOv8 dengan integrasi Global Attention Mechanism (GAM) dan GhostNet. Modifikasi ini bertujuan untuk meningkatkan kemampuan deteksi model dalam menangkap fitur penting serta mengurangi kompleksitas komputasi. Hasil optimasi menunjukkan peningkatan mean average precision (mAP) hingga 71,90% pada deteksi subtipe ALL dan efisiensi komputasi yang lebih baik dengan total waktu inferensi sebesar 7,1 ms. vi Kontribusi utama dari penelitian ini meliputi pengembangan metode data augmentasi berbasis superpixel yang efektif, penerapan metode image enhancement adaptive menggunakan CLAHE, optimasi hyperparameter dan modifikasi arsitektur YOLOv8 untuk meningkatkan akurasi dan efisiensi deteksi subtipe ALL. Penelitian ini tidak hanya meningkatkan kinerja model deteksi subtipe ALL tetapi juga mengusulkan pendekatan yang lebih efisien dan efektif dalam pengolahan serta analisis citra mikroskopis sel darah, sehingga dapat berkontribusi pada peningkatan diagnosis dan pengelolaan penyakit leukemia limfoblastik akut.
===================================================================================================================================
Microscopic image analysis of blood cells plays an important role in the diagnosis of Acute Lymphoblastic Leukemia (ALL) subtypes. Previous methods, such as Convolutional Neural networks (CNN) and traditional image processing techniques, have been used for object detection in microscopic images. However, these methods still face limitations in the integration of local and global features and are not optimal in handling visual variations in images, such as overlapping cells and dataset imbalance. Therefore, this study aims to develop a more accurate object detection method in microscopic images for ALL subtype detection by addressing several research gaps, including dataset limitations and imbalances and the integration of local and global map features. This study is divided into several main stages. The first stage focuses on the development of a superpixel-based data augmentation method to address the limitations and imbalances in microscopic image datasets of ALL subtype blood cells. This method utilizes techniques such as Gradient-weighted Class Activation Mapping (Grad-CAM), Multi-Head Attention, SaliencyMix, and OcCaMix to increase the diversity and representativeness of the data without increasing the amount of existing training data. As a result, the FocusAugMixV1 method achieved an accuracy of 99,07%, far surpassing previous methods ranging from 65% to 97,52%. The second stage involved the development of an adaptive image enhancement method using Contrast Limited Adaptive Histogram Equalization (CLAHE) to improve the quality of microscopic images by adaptively adjusting contrast and coloring, thereby reducing artifacts and process complexity. The third stage of this research focused on optimizing the hyperparameters of the YOLOv8 object detection model using the Grey Wolf Optimizer (GWO) and modifying the YOLOv8 architecture with the integration of Global Attention Mechanism (GAM) and GhostNet. This modification aims to improve the model's detection ability in capturing important features and reducing computational complexity. The optimization results showed an increase in mean Average Precision (mAP) of up to 71,90% in ALL subtype detection and better computational efficiency with a total inference time of 7,1 ms. The main contributions of this study include the development of an effective superpixel-based data augmentation method, the application of adaptive image enhancement method using CLAHE, hyperparameter optimization and modification of YOLOv8 architecture to improve the accuracy and efficiency of ALL viii subtype detection. This study not only improves the performance of the ALL subtype detection model but also proposes a more efficient and effective approach in the processing and analysis of blood cell microscopic images, which can contribute to the improvement of the diagnosis and management of acute lymphoblastic leukemia.

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	Acute Lymphoblastic Leukemia, Image Enhancement, Deep learning, Object Detection, Citra Mikroskopis Acute Lymphoblastic Leukemia, Image Enhancement, Deep learning, Object Detection, Microscopic Image
Subjects:	T Technology > T Technology (General) > T58.62 Decision support systems
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55001-(S3) PhD Thesis (Comp Science)
Depositing User:	Tanzilal Mustaqim
Date Deposited:	23 Jul 2025 01:28
Last Modified:	23 Jul 2025 01:28
URI:	http://repository.its.ac.id/id/eprint/120606

Actions (login required)

View Item