Browsing by Author "Demirci, Orhan"

Now showing 1 - 1 of 1

Tree-net: Bottleneck feature supervised network for biomedical image segmentation
(Abdullah Gül Üniversitesi, Fen Bilimleri Enstitüsü, 2024) Demirci, Orhan; AGÜ, Fen Bilimleri Enstitüsü, Elektrik ve Bilgisayar Mühendisliği Ana Bilim Dalı
In this thesis, we introduce Tree-NET, a novel approach for medical image segmentation utilizing bottleneck feature supervision. This method enhances traditional segmentation algorithms by keeping supervision between bottleneck features of the network. The primary goal is to improve the model's ability to learn discriminative and robust features while simultaneously reducing computational costs. Bottleneck feature supervision involves compressing the input and label data using Autoencoders and then supervising the bottleneck features with a segmentation network named "Bridge-Net," which can be any segmentation model of choice. We applied Tree-NET to two critical medical image segmentation tasks: skin lesion segmentation and polyp segmentation. Our experiments demonstrate significant improvements in segmentation accuracy and efficiency. For instance, the U-NET backboned Tree-NET uses only 154.43 MB for executing and storing the model, which is almost 3.5 times smaller than the original U-Net while having a close number of trainable parameters. In skin lesion segmentation, Tree-NET achieved dice, Intersection-over-Union (IoU), and accuracy scores of 0.893, 0.751, and 0.977 respectively. For polyp segmentation, the scores were 0.856, 0.795, and 0.923 for dice, IoU, and accuracy respectively. Compared to traditional segmentation models, the empirical results show that Tree-NET achieves higher accuracy with reduced training time and computational cost, thus representing a significant advancement in medical image analysis by providing more reliable and efficient tools for clinical applications.