ABELYA, RAHMA ALLYSA (2026) Bone Fracture Detection In X-Ray Images Using Detection Transformer (DETR). Undergraduate thesis, UPN Veteran Jawa Timur.
|
Text (Cover)
22081010004.-cover.pdf Download (2MB) |
|
|
Text (Bab I)
22081010004.-bab1.pdf Download (120kB) |
|
|
Text (Bab II)
22081010004.-bab2.pdf Restricted to Repository staff only until 12 November 2029. Download (575kB) |
|
|
Text (Bab III)
22081010004.-bab3.pdf Restricted to Repository staff only until 12 November 2029. Download (844kB) |
|
|
Text (Bab IV)
22081010004.-bab4.pdf Restricted to Repository staff only until 12 November 2029. Download (1MB) |
|
|
Text (Bab V)
22081010004.-bab5.pdf Download (150kB) |
|
|
Text (Daftar Pustaka)
22081010004.-daftarpustaka.pdf Download (113kB) |
Abstract
Bone fractures are common injuries, and their examination is generally performed using X-ray images. Interpreting X-ray images is not always easy because of low contrast and subtle fracture details, while the widely used CNN-based approaches still focus on classification without providing fracture location information. This study aims to apply the Detection Transformer (DETR) to detect and localize bone fractures in X-ray images using bounding boxes, and to implement it in a web-based system. The data consist of the secondary FracAtlas dataset from Kaggle, containing 717 fracture images and 717 non-fracture images. The research stages include literature study, dataset collection and analysis, pre-processing (resizing, normalization, tensor conversion, and padding), training the DETR model with a ResNet-50 backbone in eight configuration scenarios, and evaluation using Average Precision (AP) and Average Recall (AR). The best configuration was obtained with a 70%-15%-15% data split, a learning rate of 1×10⁻⁴, and 300 epochs, yielding an AP@[0.50:0.95] of 0.162, an AP@0.50 of 0.396, and an AR of 0.338. At a threshold of 0.5, the model achieved a sensitivity of 0.6606, a specificity of 0.9083, and an accuracy of 0.7844. These results show that DETR is able to produce fracture detection predictions, although localization accuracy at higher IoU remains limited. The web-based system built with Streamlit successfully functions as a tool for visualizing detection results.
| Item Type: | Thesis (Undergraduate) | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Contributors: |
|
||||||||||||
| Subjects: | T Technology > T Technology (General) > T58.6-58.62 Management Information Systems | ||||||||||||
| Divisions: | Faculty of Computer Science > Departemen of Informatics | ||||||||||||
| Depositing User: | Rahma Allysa Abelya | ||||||||||||
| Date Deposited: | 12 Nov 2026 03:13 | ||||||||||||
| Last Modified: | 12 Nov 2026 04:19 | ||||||||||||
| URI: | https://repository.upnjatim.ac.id/id/eprint/61258 |
Actions (login required)
![]() |
View Item |
