<xml> </xml><![endif]--><!--[if gte mso 9]><xml> Normal 0 false false false EN-ID JA X-NONE </xml><![endif]--><!--[if gte mso 9]><xml> </xml><![endif]--><!--[if gte mso 10]> <style> /* Style Definitions */ table.MsoNormalTable {mso-style-name:"Tabel Normal"; mso-tstyle-rowband-size:0; mso-tstyle-colband-size:0; mso-style-noshow:yes; mso-style-priority:99; mso-style-parent:""; mso-padding-alt:0in 5.4pt 0in 5.4pt; mso-para-margin-top:6.0pt; mso-para-margin-right:0in; mso-para-margin-bottom:6.0pt; mso-para-margin-left:0in; text-align:justify; text-justify:inter-ideograph; line-height:150%; mso-pagination:widow-orphan; font-size:11.0pt; font-family:"Calibri",sans-serif; mso-ascii-font-family:Calibri; mso-ascii-theme-font:minor-latin; mso-hansi-font-family:Calibri; mso-hansi-theme-font:minor-latin; mso-bidi-font-family:"Times New Roman"; mso-bidi-theme-font:minor-bidi;} </style> <![endif]-->Manual diagnosis of Acute Myeloid Leukemia (AML) through microscopic analysis of blood smears is a time-consuming, subjective process that requires substantial resources. To address these challenges, deep learning approaches such as U-Net have been adopted due to their high accuracy in medical image segmentation. However, the computational complexity of such models poses a significant barrier to their implementation on portable embedded systems with limited power and computing resources. This study aims to develop an efficient, real-time white blood cell segmentation system using a modified U-Net architecture. The modification involves replacing U-Net’s standard encoder with more efficient modern backbones, specifically EfficientNet-B4 and RepViT-M1.5. The models were trained on a combined dataset of public and clinical microscopic images and evaluated on the NVIDIA Jetson Orin Single Board Computer using metrics such as Intersection over Union (IoU), precision, recall, and inference speed (FPS). The results demonstrate that converting the models to the TensorRT FP16 format significantly boosts inference speed, from around 4–7 FPS to over 45 FPS. Among the tested backbones, RepViT-M1.5 achieved the best overall performance, with a mean IoU exceeding 95% and superior real-time visual segmentation quality. This study confirms that a U-Net model with the RepViT-M1.5 backbone, optimized with TensorRT, offers an effective and efficient solution for real-time AML cell segmentation on edge devices."> <xml> </xml><![endif]--><!--[if gte mso 9]><xml> Normal 0 false false false EN-ID JA X-NONE </xml><![endif]--><!--[if gte mso 9]><xml> </xml><![endif]--><!--[if gte mso 10]> <style> /* Style Definitions */ table.MsoNormalTable {mso-style-name:"Tabel Normal"; mso-tstyle-rowband-size:0; mso-tstyle-colband-size:0; mso-style-noshow:yes; mso-style-priority:99; mso-style-parent:""; mso-padding-alt:0in 5.4pt 0in 5.4pt; mso-para-margin-top:6.0pt; mso-para-margin-right:0in; mso-para-margin-bottom:6.0pt; mso-para-margin-left:0in; text-align:justify; text-justify:inter-ideograph; line-height:150%; mso-pagination:widow-orphan; font-size:11.0pt; font-family:"Calibri",sans-serif; mso-ascii-font-family:Calibri; mso-ascii-theme-font:minor-latin; mso-hansi-font-family:Calibri; mso-hansi-theme-font:minor-latin; mso-bidi-font-family:"Times New Roman"; mso-bidi-theme-font:minor-bidi;} </style> <![endif]-->Manual diagnosis of Acute Myeloid Leukemia (AML) through microscopic analysis of blood smears is a time-consuming, subjective process that requires substantial resources. To address these challenges, deep learning approaches such as U-Net have been adopted due to their high accuracy in medical image segmentation. However, the computational complexity of such models poses a significant barrier to their implementation on portable embedded systems with limited power and computing resources. This study aims to develop an efficient, real-time white blood cell segmentation system using a modified U-Net architecture. The modification involves replacing U-Net’s standard encoder with more efficient modern backbones, specifically EfficientNet-B4 and RepViT-M1.5. The models were trained on a combined dataset of public and clinical microscopic images and evaluated on the NVIDIA Jetson Orin Single Board Computer using metrics such as Intersection over Union (IoU), precision, recall, and inference speed (FPS). The results demonstrate that converting the models to the TensorRT FP16 format significantly boosts inference speed, from around 4–7 FPS to over 45 FPS. Among the tested backbones, RepViT-M1.5 achieved the best overall performance, with a mean IoU exceeding 95% and superior real-time visual segmentation quality. This study confirms that a U-Net model with the RepViT-M1.5 backbone, optimized with TensorRT, offers an effective and efficient solution for real-time AML cell segmentation on edge devices.">
Laporkan Masalah

SEGMENTASI SEL DARAH PUTIH MENGGUNAKAN MODIFIKASI U-NET PADA EMBEDDED ARTIFICIAL INTELLIGENCE

Wisnu Aryo Jatmiko, Ika Candradewi, S.Si., M.Cs.

2025 | Skripsi | ELEKTRONIKA DAN INSTRUMENTASI

<!--[if gte mso 9]><xml> </xml><![endif]--><!--[if gte mso 9]><xml> Normal 0 false false false EN-ID JA X-NONE </xml><![endif]--><!--[if gte mso 9]><xml> </xml><![endif]--><!--[if gte mso 10]> <style> /* Style Definitions */ table.MsoNormalTable {mso-style-name:"Tabel Normal"; mso-tstyle-rowband-size:0; mso-tstyle-colband-size:0; mso-style-noshow:yes; mso-style-priority:99; mso-style-parent:""; mso-padding-alt:0in 5.4pt 0in 5.4pt; mso-para-margin-top:6.0pt; mso-para-margin-right:0in; mso-para-margin-bottom:6.0pt; mso-para-margin-left:0in; text-align:justify; text-justify:inter-ideograph; line-height:150%; mso-pagination:widow-orphan; font-size:11.0pt; font-family:"Calibri",sans-serif; mso-ascii-font-family:Calibri; mso-ascii-theme-font:minor-latin; mso-hansi-font-family:Calibri; mso-hansi-theme-font:minor-latin; mso-bidi-font-family:"Times New Roman"; mso-bidi-theme-font:minor-bidi;} </style> <![endif]-->Diagnosis manual Acute Myeloid Leukemia (AML) melalui analisis mikroskopis apusan darah merupakan proses yang memakan waktu, bersifat subjektif, dan membutuhkan banyak sumber daya. Untuk mengatasi tantangan ini, pendekatan berbasis deep learning seperti U-Net telah digunakan karena kemampuannya dalam memberikan akurasi tinggi dalam segmentasi citra medis. Namun, kompleksitas komputasi dari model ini menjadi hambatan dalam implementasinya pada sistem tertanam portabel yang memiliki keterbatasan daya dan sumber daya komputasi.

Penelitian ini bertujuan mengembangkan sistem segmentasi sel darah putih yang efisien dan real-time dengan menggunakan arsitektur U-Net yang dimodifikasi. Modifikasi dilakukan dengan mengganti encoder standar U-Net dengan backbone modern yang lebih efisien, yaitu EfficientNet-B4 dan RepViT-M1.5. Model dilatih menggunakan kombinasi dataset citra mikroskopis dari sumber publik dan klinis, dan dievaluasi pada perangkat Single Board Computer NVIDIA Jetson Orin menggunakan metrik Intersection over Union (IoU), presisi, recall, serta kecepatan inferensi (FPS).

Hasil penelitian menunjukkan bahwa konversi model ke format TensorRT FP16 memberikan peningkatan signifikan terhadap kecepatan inferensi, dari 4–7 FPS menjadi lebih dari 45 FPS. Di antara kedua backbone yang diuji, RepViT-M1.5 memberikan performa terbaik dengan rata-rata IoU diatas 95% serta segmentasi visual real-time yang unggul. Penelitian ini membuktikan bahwa kombinasi U-Net dengan RepViT-M1.5 dan optimisasi TensorRT merupakan solusi efektif dan efisien untuk segmentasi sel AML secara real-time pada perangkat edge.

<!--[if gte mso 9]><xml> </xml><![endif]--><!--[if gte mso 9]><xml> Normal 0 false false false EN-ID JA X-NONE </xml><![endif]--><!--[if gte mso 9]><xml> </xml><![endif]--><!--[if gte mso 10]> <style> /* Style Definitions */ table.MsoNormalTable {mso-style-name:"Tabel Normal"; mso-tstyle-rowband-size:0; mso-tstyle-colband-size:0; mso-style-noshow:yes; mso-style-priority:99; mso-style-parent:""; mso-padding-alt:0in 5.4pt 0in 5.4pt; mso-para-margin-top:6.0pt; mso-para-margin-right:0in; mso-para-margin-bottom:6.0pt; mso-para-margin-left:0in; text-align:justify; text-justify:inter-ideograph; line-height:150%; mso-pagination:widow-orphan; font-size:11.0pt; font-family:"Calibri",sans-serif; mso-ascii-font-family:Calibri; mso-ascii-theme-font:minor-latin; mso-hansi-font-family:Calibri; mso-hansi-theme-font:minor-latin; mso-bidi-font-family:"Times New Roman"; mso-bidi-theme-font:minor-bidi;} </style> <![endif]-->Manual diagnosis of Acute Myeloid Leukemia (AML) through microscopic analysis of blood smears is a time-consuming, subjective process that requires substantial resources. To address these challenges, deep learning approaches such as U-Net have been adopted due to their high accuracy in medical image segmentation. However, the computational complexity of such models poses a significant barrier to their implementation on portable embedded systems with limited power and computing resources.

This study aims to develop an efficient, real-time white blood cell segmentation system using a modified U-Net architecture. The modification involves replacing U-Net’s standard encoder with more efficient modern backbones, specifically EfficientNet-B4 and RepViT-M1.5. The models were trained on a combined dataset of public and clinical microscopic images and evaluated on the NVIDIA Jetson Orin Single Board Computer using metrics such as Intersection over Union (IoU), precision, recall, and inference speed (FPS).

The results demonstrate that converting the models to the TensorRT FP16 format significantly boosts inference speed, from around 4–7 FPS to over 45 FPS. Among the tested backbones, RepViT-M1.5 achieved the best overall performance, with a mean IoU exceeding 95% and superior real-time visual segmentation quality. This study confirms that a U-Net model with the RepViT-M1.5 backbone, optimized with TensorRT, offers an effective and efficient solution for real-time AML cell segmentation on edge devices.

Kata Kunci : Sel Darah Putih, Segmentasi, U-Net, Sistem Tertanam, TensorRT.

  1. S1-2025-482380-abstract.pdf  
  2. S1-2025-482380-bibliography.pdf  
  3. S1-2025-482380-tableofcontent.pdf  
  4. S1-2025-482380-title.pdf