Feasibility Study of Deep Learning Tumor Segmentation for a Merged Tumor Dataset: Head & Neck and Limbs
- PDF / 658,917 Bytes
- 6 Pages / 595.22 x 842 pts (A4) Page_size
- 95 Downloads / 154 Views
Feasibility Study of Deep Learning Tumor Segmentation for a Merged Tumor Dataset: Head & Neck and Limbs Ye-In Park, Sang-Won Kang, Kyeong-Hyeon Kim and Tae Suk Suh∗ Department of Biomedical Engineering, The Catholic University of Korea, Seoul 06591, Korea
Jin-Beom Chung Department of Radiation Oncology, Seoul National University Bundang Hospital, Seongnam-si 13620, Korea (Received 3 July 2020; revised 2 September 2020; accepted 11 September 2020) The aim of this study is to evaluate the feasibility of a deep learning tumor segmentation network trained by merged tumor dataset. PET-CT datasets for head-and-neck (H&N) and limb tumors were used to train three different networks: H&N, Limb, and merged (H&N + Limb). Dice similarity coefficient (DSC) of the merged network (0.89) in limb tumors was the same as that of the Limb network. In H&N tumor, DSC of the merged network (0.72) was higher than that of the H&N network (0.69). We found that the merged network could be applied simultaneously in H&N and limb tumor segmentation. Keywords: Deep learning (DL), Convolutional neural network (CNN), Tumor segmentation DOI: 10.3938/jkps.77.1049
I. INTRODUCTION Tumor segmentation is very important in treatment planning of radiation therapy. Inaccurate segmentation can lead to an overdose to normal organs and an underdose in the gross tumor volume (GTV) [1, 2]. Positron emission tomography (PET) and other functional information have been applied to improve the segmentation accuracy for the GTV. Despite the use of PET, tumor segmentation still inaccurate when performed manually by radiation oncologists [3]. Observer experience and protocol differences in manual contouring derive variations of segmentation for GTVs [4]. Therefore, automatic tumor segmentation is required to reduce the variability of manual contouring. Various automatic segmentation methods have been applied to obtain fast and highly reproducible results in GTV contouring. The threshold [5] and the region growing [6] methods provide simple GTVs to be segmented according to the intensity and the gradient of medical images. Machine learning (ML) based segmentations, such as the Markov random field [7], fuzzy c-means [8], and Bayesian algorithm [9] have been applied to GTV segmentation. These methods have provided more accurate GTVs with smoother boundaries than the threshold method. However, previous studies have reported that several ML methods have limited performance in complicated GTV cases such as a diffuse or multi-focal tumor ∗ E-mail:
[10]. Recently, deep learning has been applied to automatic tumor segmentation studies to take into account the features of GTVs, which are more complex than that of ML methods. A convolutional neural network (CNN), described by Huang et al [11], achieved a dice similarity coefficient (DSC) similar to that of Markov random fields. Xu et al [12] reported that V-net tumor contours had better specificity and precision scores than the random forest, k-nearest neighbor, and support vector machine algorithms. Various CNN structures have been appli
Data Loading...
