Nutritional monitoring in Indonesia still relies on manual 24-hour dietary recall , which is prone to recall bias, inaccurate portion estimation, and subjectivity. This limitation directly impacts early detection of stunting and malnutrition , especially in community-based healthcare settings such as posyandu.
Syahreza Adnan Al Azhar (Team Lead)
Led a cross-functional team to design and deliver the TumbuhSehat application
Authored the funded proposal and final report
Designed and implemented the mobile application (Flutter, Clean architecture)
Conducted model experiments for semantic segmentation of food (Dense UNet, DeepLabV3, YOLOv11 seg)
Designed an end-to-end nutrition calculation workflow aligned with Ministry of Health & WHO standards
Contributed to the UI UX design of the TumbuhSehat application
Puri Lalita Anagata — Managed project timeline and task; contributed to the UI UX design; contributed to writing the proposal and final report
Muhammad Fajar Mufid — Designed and developed backend service
Yolanda Rahmah Christy — Contributed to the UI UX design; designed the social media of TumbuhSehat
Valentino Hartanto — Contributed to writing the proposal and final report; conducted model experiments for depth estimation (MiDaS, DPT, Open3D, ColMap)
The system combines semantic segmentation and monocular depth estimation to estimate food volume from a single top-down image . Segmentation masks are fused with calibrated depth maps to compute food volume, which is then mapped to nutritional values using Ministry of Health Standards, and WHO Child Growth Standards. The applied Ministry of Health standards include the Nutritional Adequacy Rate (Angka Kecukupan Gizi - AKG), the Indonesian Food Composition Table (Tabel Komposisi Pangan Indonesia - TKPI), and the Food Exchange List (Daftar Bahan Makanan Penukar).
- Chose DeepLabV3 for stable boundary segmentation on food textures
- Used MiDaS v3.1 to avoid hardware dependency (no depth sensor)
Applied plate-diameter calibration (25 cm) to convert relative depth into metric volume
Designed system to work offline-first for low-connectivity environments
Model Performance
- Segmentation mIoU: 93%
- Depth estimation MRE: 4%, AbsRel: 6%
User Validation
- SUS Score: 85.49 (Excellent)
- Tested in real posyandu environments
- Validated by certified nutritionist
Instagram Link TumbuhSehat's Social Media
Organization Repository Link TumbuhSehat's Repository
File Link Intellectual Property Rights of TumbuhSehat Application
- Fully functional mobile application (Android & iOS)
- Registered Intellectual Property (HKI)
Scientific article prepared (IEEE format, planned Scopus Q3 submission)
- Potential continuation to PKM-KI & public health deployment
The main technical limitation lies in handling mixed and translucent food textures under uneven lighting. If revisited, I would explore lightweight attention mechanisms and larger-scale dataset expansion to improve generalization across food categories.