Aplikasi Pengukuran Antropometri Lingkar Lengan Atas (LILA) untuk Mengidentifikasi Kemungkinan Risiko Malnutrisi Menggunakan CNN dan Regresi Linear

Abstract

Mid-Upper Arm Circumference (MUAC) measurement is a crucial indicator for the early identification of malnutrition risk. Despite its importance, traditional manual measurement methods face severe reliability issues, characterized by high inter-observer variability. Furthermore, a lack of public understanding regarding the importance of MUAC and its health implications results in low awareness and insufficient health monitoring. Therefore, there is a critical need to develop an automated Deep Learning-based solution to establish precision and reliability that surpasses manual performance. This study aims to develop an automated non-contact MUAC measurement system utilizing a Convolutional Neural Network with U-Net architecture for arm segmentation and linear regression techniques for 3D circumference estimation, thereby improving identification effectiveness and accuracy. The system is designed as a multitask problem integrating segmentation, calibration, and regression. The CNN, utilizing the U-Net architecture, serves as the backbone to accurately identify and isolate upper arm pixels from input images. For accurate circumference estimation, linear regression is applied to 2D distance measurements from images captured from multiple angles (multi-view), specifically front and side views. The linear regression approach proved superior to methods based on rigid geometric models. The performance results of the U-Net model for image segmentation are quite good with a Mean Precision value of 0.9629; Mean Recall of 0.9284; Mean F1-Score of 0.9448; and Mean IoU value of 0.8988. The developed Linear Regression Model has quite good performance in predicting MUAC size with Mean Absolute Error (MAE) results of 1,12 cm; Mean Percentage Error of 4,19% and Root Mean Square Error (RMSE) of 1,39 cm.