Perhitungan Jumlah Sel Pada Citra Hematoksilin Eosin Menggunakan Deep Learning

Arjuna, Dzaky Hanif (2026) Perhitungan Jumlah Sel Pada Citra Hematoksilin Eosin Menggunakan Deep Learning. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6022241128-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (2MB) | Request a copy

Abstract

Diagnosis kanker payudara melalui analisis manual citra histopatologi Hematoksilin Eosin (HE) masih menghadapi kendala berupa waktu yang lama dan tingkat subjektivitas yang tinggi. Penelitian ini bertujuan mengembangkan sistem otomatis berbasis deep learning untuk penghitungan jumlah sel kanker payudara secara otomatis dengan membandingkan dua pendekatan, yaitu segmentasi multi-class dan segmentasi dengan klasifikasi. Arsitektur U-Net digunakan untuk segmentasi, serta MobileNetV2 dan VGG19 untuk klasifikasi, yang diuji pada dataset publik IHC4BC dan data klinis dari sebuah rumah sakit di Jawa Timur. Penelitian ini juga mengevaluasi dampak strategi pengolahan data Resizing dibandingkan Tiling terhadap performa model. Hasil eksperimen pada dataset IHC4BC menunjukkan bahwa model klasifikasi MobileNetV2 dan VGG19 mencapai akurasi pengujian sebesar 98,80%. Namun, pada sistem dua tahap, evaluasi perhitungan sel (counting) menunjukkan kesalahan yang tinggi dengan Mean Absolute Error (MAE) sel positif sebesar 119,87, yang disebabkan oleh kegagalan pemisahan sel yang saling berhimpitan pada tahap segmentasi biner. Sebaliknya, pendekatan segmentasi multi-class dengan strategi tiling (Skenario 5) terbukti jauh lebih unggul, dengan MAE sel positif terendah sebesar 18,46 dan MAE sel negatif sebesar 1,66. Pengujian pada data klinis rumah sakit di Jawa Timur mengonfirmasi keunggulan pendekatan ini, di mana segmentasi multi-class menghasilkan MAE positif 44,40, jauh lebih baik dibandingkan pendekatan dua tahap yang mencapai MAE positif 123,35. Penelitian ini menyimpulkan bahwa metode segmentasi multi-class dengan strategi tiling merupakan pendekatan yang paling efektif untuk penghitungan sel otomatis pada citra histopatologi kanker payudara, dengan MAE penghitungan terendah pada data uji IHC4BC (positif 18,46).
=====================================================================================================================================
Manual analysis of Hematoxylin and Eosin (HE) histopathological images for breast cancer diagnosis still faces significant challenges, including timeconsuming processes and high subjectivity. This study aims to develop an automated system based on Deep Learning for automated counting of breast cancer cells by comparing two approaches: Multi-class Segmentation and Segmentation with Classification. The U-Net architecture is used for segmentation, while MobileNetV2 and VGG19 are employed for classification, both evaluated on the public IHC4BC dataset and clinical data from a hospital in East Java. This study also evaluates the impact of Resizing and Tiling data preprocessing strategies on model performance. Experimental results on the IHC4BC dataset demonstrate that both MobileNetV2 and VGG19 classification models achieve a test accuracy of 98.80%. However, in the two-stage pipeline, cell counting evaluation reveals high errors with a Mean Absolute Error (MAE) of 119.87 for positive cells, attributed to the failure of the binary segmentation stage in separating touching cells. In contrast, the Multi-class Segmentation approach with the Tiling strategy (Scenario 5) proved significantly superior, achieving the lowest positive cell MAE of 18.46 and a negative cell MAE of 1.66. Validation on clinical data from the hospital in East Java confirms this superiority, where Multi-class Segmentation yielded a positive MAE of 44.40, substantially better than the two-stage approach which reached a positive MAE of 123.35. This study concludes that the Multi-class Segmentation method with the Tiling strategy is the most effective approach for automated cell counting in breast cancer histopathological images, achieving the lowest counting MAE on the IHC4BC test set (positive 18.46).

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Convolutional Neural Network, deep learning, hematoksilineosin, kanker payudara, klasifikasi sel Breast Cancer, Cell Classification, Convolutional Neural Network, Deep Learning, Hematoxylin and Eosin
Subjects:	T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques. Image analysis--Data processing.
Divisions:	Faculty of Electrical Technology > Electrical Engineering > 20101-(S2) Master Thesis
Depositing User:	Dzaky Hanif Arjuna
Date Deposited:	30 Jul 2026 02:34
Last Modified:	30 Jul 2026 02:34
URI:	http://repository.its.ac.id/id/eprint/139988

Actions (login required)

View Item