Klasifikasi Penyakit Acute Lymphoblastic Leukemia Menggunakan Convolutional Neural Network Efficientnet B4 Dan B7

Kusumastuti , Rajnaparamitha
Wariyanto Abdullah, Robi
Handoko

Jurnal Teknologi Informasi dan Ilmu Komputer
Vol 13 No 4: Agustus 2026

Leukemia Limfoblastik Akut (ALL) merupakan leukemia paling umum pada anak-anak sehingga memerlukan deteksi dini berbasis analisis citra yang cepat dan akurat. Penelitian ini mengklasifikasikan citra apusan darah tepi (Peripheral Blood Smear/PBS) menggunakan Convolutional Neural Network (CNN) berarsitektur EfficientNet B4 dan EfficientNet B7. Dataset terdiri atas 5.733 citra dari 89 pasien, diproses pada ukuran input 224 × 224 piksel, kemudian dibagi menjadi data latih, validasi, dan uji dengan rasio 80:10:10, dengan jumlah 4.586 data pelatihan, 573 data validasi, dan 574 data uji. Pelatihan dilakukan dengan pendekatan transfer learning dan augmentasi data untuk meningkatkan kemampuan generalisasi model. Evaluasi dilakukan menggunakan akurasi, precision, recall, F1-score, training loss, dan validation loss. Hasil terbaik menunjukkan bahwa EfficientNet B4 pada skenario 100 epoch mencapai akurasi 99,50%, precision 98,41%, recall 98,43%, dan F1-score 99,20%. Model ini mampu memberikan performa kompetitif terhadap EfficientNet B7 dengan kebutuhan komputasi lebih rendah serta stabilitas lebih baik terhadap overfitting. Kontribusi penelitian ini terletak pada analisis trade-off antara akurasi, stabilitas validasi, risiko overfitting, dan efisiensi komputasi model sebagai dasar pemilihan arsitektur yang lebih tepat untuk deteksi dini leukemia.   Abstract Acute Lymphoblastic Leukemia (ALL) is the most common type of leukemia in children, requiring early detection through fast and accurate image-based analysis. This study classifies peripheral blood smear (PBS) images using a Convolutional Neural Network (CNN) with EfficientNet B4 and EfficientNet B7 architectures. The dataset consists of 5,733 images from 89 patients, processed at an input size of 224 × 224 pixels, and divided into training, validation, and testing sets with an 80:10:10 ratio, comprising 4,586 training images, 573 validation images, and 574 testing images. Training was conducted using transfer learning and data augmentation to improve the model’s generalization ability. Evaluation was performed using accuracy, precision, recall, F1-score, training loss, and validation loss. The best result was achieved by EfficientNet B4 in the 100-epoch scenario, with an accuracy of 99.50%, precision of 98.41%, recall of 98.43%, and F1-score of 99.20%. This model provides competitive performance compared with EfficientNet B7 while requiring lower computational resources and demonstrating better stability against overfitting. The contribution of this study lies in analyzing the trade-off between accuracy, validation stability, overfitting risk, and computational efficiency as a basis for selecting a more appropriate architecture for early leukemia detection.