Evidence record · UDINUS · SDGs
- Unit
- Rektorat UDINUS
- Period
- 2024
- Goals
- Record
- a9bfca64-1f87-414a-8b9a-188ca14f5d32
- Published
- 2026-10-30
Universitas Dian Nuswantoro presented research on a framework for low-light image segmentation of tuberculosis bacilli using a semisupervised approach. The activity involved UDINUS researchers Wulandari, Sari Ayu; Eddy Purnama, I Ketut; Yuniarno, Eko Mulyanto; and Purnomo, Mauridhi Hery. The research addressed the need for annotated images as ground truth for deep learning and the time and effort required for manual bacterial segmentation. The activity also addressed the challenge presented by low-light bacterial images, which affect automatic annotation tasks.
The proposed system used a fuzzy-based clustering method to improve bacterial object segmentation through a multicluster concept. The State Transform Algorithm was used to obtain starting centroids because Kernel Intuitionistic Fuzzy C-Means is sensitive to starting centroids and can become stuck in local optima. The research reported that KIFCM-STA with bicluster produced poor accuracy for low-light images. As a continuation hybrid of KIFCM-STA, the multicluster technique, identified as MKIFCM-STA, was presented to improve performance.
The framework determined an ideal number of clusters using the Silhouette method and applied cluster ranking with Topsis to identify clusters containing bacterial objects. The activity compared the proposed method with four prior approaches: IFCM, KFCM, KIFCM, and KIFCM-STA. The comparison considered qualitative aspects through image visualization and quantitative aspects through average IOU, Dice, HD, ASD, and Accuracy. In low-light image clustering tasks, the model achieved an accuracy score of 89.438% and was reported to significantly improve performance. These findings relate the activity to the stated goal through research addressing image-based work related to tuberculosis bacilli and bacterial segmentation. Through this research activity, UDINUS presented a framework focused on low-light bacterial image clustering and the improvement of image auto-annotation procedures.