Deep Learning-based Automated Delineation of Head and Neck Malignant Lesions from PET Images

Deep Learning-based Automated Delineation of Head and Neck Malignant Lesions from PET Images
2020 IEEE Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC)
（ Conference papers , Article only 3 page , Less content , But data sets can be referred to ）

Abstract

Accurately describe the gross tumor volume (GTV) Is the key to radiation oncology treatment planning . Because the shape of malignant lesions is irregular and diverse , This task is very challenging . stay PET Hand painted on the image GTV Not only time-consuming , It is also affected by variability between and within observers . In this work , We have developed a method based on deep learning , For use in patients with head and neck cancer PET Automatically draw on the image GTV. So , A full convolution neural network for volume medical image segmentation is adopted VNet and 20 Layer residual convolution neural network HighResNet. stay 510 Cases of head and neck cancer 18F-FDG-PET/CT On the image , Use manual definitions ( Reference resources ) Of GTV Train these algorithms 、 Evaluation and testing . The input of these networks ( Whether in the training phase or in the evaluation phase ) All are 12×12×12 cm Of PET The sub volume of an image , Contains the full volume of the tumor and the adjacent background radiotracer uptake . These networks are trained to generate representative inputs PET On a subvolume GTV Binary mask for . Standard segmentation metrics , Include Dice Similarity and precision It is used to evaluate the performance of these algorithms .HighResNet Automatic GTV describe ,Dice Index is 0.87±0.04, and V-net by 0.86±0.06. Although the performance of the two methods is similar , but HighResNet There was little difference between different subjects , Reflected in small standard deviation and significantly higher precision Index (0.87±0.07 vs 0.80±0.10). Deep learning techniques , especially HighResNet Algorithm , In the head and neck PET Automatic of images GTV The sketch shows good performance . Combined with anatomy / structural information , especially MRI, It may lead to higher segmentation accuracy or less difference between different objects .

Keywords

Head and Neck Cancer Segmentation PET Deep Learning

Introduction

Head and neck cancer is one of the most common types of cancer in the world . High precision external radiotherapy (RT) It is considered to be an effective strategy to treat this kind of cancer . Accurately delineate the gross tumor volume (GTV) Is to carry out effective image guidance RT Key steps in planning .
Accurate and robust GTV Depiction allows the maximum therapeutic dose to be delivered to the target volume and protects healthy tissues from unwanted toxicity . Due to its proximity to key anatomical structures , This problem is particularly important in head and neck tumors . In clinical practice ,GTV Delineation is usually performed by radiation oncologists through manual or semi manual techniques [1]. This process is not only time consuming , And it is easy to see the difference between the inside of the observer and the outside of the observer . Many studies show that , Due to the knowledge of the operator , Experience and workload ,GTV The multi - observer descriptions vary greatly [2]. For this reason , accuracy , Robust and repeatable GTV The depiction method is effective for head and neck tumors RT The plan is essential .
Because of the irregular shape of the tumors in the head and neck 、 Proximity to key structures and presence appearance / Healthy tissue with similar signals , So the head and neck tumor GTV Segmentation is challenging [36]. The method based on deep learning shows good performance in medical image analysis , Including image artifact removal [7,8]、 Image enhancement [9-11]、 quantitative [12] And segmentation [13,14]. In this work , We have developed a method based on deep learning , From the head and neck PET Auto sketch in image GTV. So , A full convolution neural network for volume medical image segmentation is adopted V-Net and 20 Layer residual convolution neural network HighResNet Algorithm . The purpose of this study is to test the segmentation accuracy of deep learning methods in different disciplines 、 The level of overall performance and variability .

Materials and methods

Description of PET/CT datasets

The dataset includes 510 Of patients with head and neck cancer 18F-FDGPET/CT Images , These images come from Cancer imaging archives (TCIA)[15]. By experienced nuclear medicine doctors in PET Manually draw the corresponding... On the image GTV, And saved in binary format . In order to PET Images are provided to the deep learning algorithm , From primitive PET The whole tumor volume and the radioactive tracer uptake in the adjacent background are extracted from the image 12×12×12 cm Subvolumetric PET Images , To create a dataset with consistent matrix and voxel sizes . Deal with representatives accordingly GTV Binary mask for .

Deep learning approaches

In this research, two most advanced deep learning methods have been realized (V-Net and HighResNet), They have good performance in medical image segmentation [16,17].V-Net It is a full convolution neural network for volume medical image segmentation , It consists of 10 Level composition , Different feelings from 5×5×5 To 551×551×551 Unequal .HighResNet Is a convolution operation equipped with expansion Bold style Of 20 Layer residual convolutional neural network . These networks take the form of Dice The exponent is the loss function , Batch size is 1, use 10 Second cross validation program for training . With SUV Scaled in units PET Crop the image as input to these algorithms , Generate corresponding GTV Binary mask as output .

Clinical evaluation

By manual segmentation GTVs For reference , Assessed V-Net and HiResNet The performance of the model . So ,Dice indicators (Eq.1)、Jaccard Similarity coefficient (Eq.2)、precision(Eq.3), Matthew correlation coefficient (MCC)(Eq.4) Yes 510 Subjects were calculated .

Results and Discussion

chart 1 Describes the use of V-Net and HighResNet Algorithm in PET Depict... On an image GTV A representative example of . A 3D rendered volume of the tumor is also provided for visual inspection .

chart 1. utilize V-Net and HighResNet Algorithm in PET Draw on the image GTV A typical example of .
chart 2 Show for HighResNet and V-Net Algorithmic 510 Box graph of segmentation metrics for subjects , Include Dice、Jaccard、Precision and MCC. stay Dice In terms of index , Both of the two deep learning algorithms show good performance . However ,HighResNet Your performance is slightly better than V-Net,Dice by 0.87±0.04, and V-Net by 0.86±0.06. Besides ,HighResNet It showed less difference and stronger robustness among different subjects , This is reflected in the lower standard deviation .

chart 2.ResNet and vNet Box diagram of different parameters
surface 1 Sum up HighResNet and V-Net Statistical summary of the segmentation metrics of the algorithm . stay precision Measurement ,HighResNet The algorithm performs better than V-Net The performance of the model .HighResNet Algorithm precision by 0.87±0.07, and V-net Model precision by 0.80±0.10. stay HighResNet In the model , Not only 510 Total of subjects precision Higher , The lower standard deviation proves that the model is suitable for automatic GTV The robustness of the description .

Conclusion

Deep learning techniques , especially HighResNet Algorithm , In the head and neck PET Automatic of images GTV The sketch shows good performance .HighResNet The algorithm not only makes the quantification of lesion volume slightly improved , And it showed little variability among different subjects . Combined with anatomy / structural information , especially MRI, It may lead to higher segmentation accuracy or less variability between different objects .