Exploring the use of efficient deep learning algorithms for lower grade gliomas cancer MRI image segmentation: A case study

Document Type : Original Article

Authors

1 School of Mechanical Engineering, College of Engineering, University of Tehran, Tehran, Iran

2 School of Computer Engineering, College of Engineering, University of Shiraz, Shiraz, Iran

Abstract

This paper presents a study on the use of efficient deep learning algorithms for lower grade gliomas (LGG) cancer image segmentation. The study compares the performance of various pretrained atrous-convolutional architectures and pre-trained U-Nets, and proposes a transformer-based approach for fast and efficient LGG segmentation. The study evaluated the performance of various models and found that DeepLabV3+ with MobileNetV3 as a backbone achieved the best per-formance among the pretrained models. However, the proposed transformer-based approach surpassed the aforementioned methods and achieved competitive results with higher scores. The study also employed transfer learning techniques to fine-tune the pretrained models on the LGG dataset, which significantly im-proved segmentation performance with a relatively low amount of training samples. The study highlights the importance of selecting the appropriate pre-trained model for the specific segmentation task. The proposed transformer-based approach offers several advantages over traditional convolutional neural networks, including efficient use of memory and better generalization. It can also process images of arbitrary sizes, making it more flexible and scalable for use in clinical settings. Segmenting medical images presents a difficult task because of the complexity of medical images and variations in imaging conditions. The use of efficient deep learning algorithms can help address these challenges by reducing computational cost, training time, and improving segmentation performance. The findings of this study can be useful in the development of accurate and efficient diagnostic tools for LGG cancer detection and treatment planning. The proposed transformer-based approach has the potential to improve medical image segmentation for other types of cancers and diseases. Overall, this study demonstrates the potential impact of deep learning and transfer learning application in medical image segmentation, with significant implications for improving cancer diagnosis and treatment.

Keywords