📝 Abstract

Pseudoprogression (PsP) is an imitation of true tumor progression (TTP) in patients with glioblastoma multiform (GBM). Differentiating them is a challenging and time-consuming task for radiologists. Although deep neural networks can automatically diagnose PsP and TTP, interpretability shortage is always the heel of Achilles. To overcome these shortcomings and win the trust of physician, we propose a transparency guided ensemble convolutional neural network to automatically stratify PsP and TTP on magnetic resonance imaging (MRI). A total of 84 patients with GBM are enrolled in the study. First, three typical convolutional neutral networks (CNNs) – VGG, ResNet and DenseNet – are trained to distinguish PsP and TTP on the dataset. Subsequently, we use the class-specific gradient information from convolutional layers to highlight the important regions in MRI. Radiological experts are then recruited to select the most lesion-relevant layer of each CNN. Finally, the selected layers are utilized to guide the construction of multi-scale ensemble CNN. The classified accuracy of the presented network is 90.20%, the promotion of specificity reaches more than 20%. The results demonstrate that network transparency and ensemble can enhance the reliability and accuracy of CNNs. The presented network is promising for the diagnosis of PsP and TTP.

💡 Analysis

Pseudoprogression (PsP) is an imitation of true tumor progression (TTP) in patients with glioblastoma multiform (GBM). Differentiating them is a challenging and time-consuming task for radiologists. Although deep neural networks can automatically diagnose PsP and TTP, interpretability shortage is always the heel of Achilles. To overcome these shortcomings and win the trust of physician, we propose a transparency guided ensemble convolutional neural network to automatically stratify PsP and TTP on magnetic resonance imaging (MRI). A total of 84 patients with GBM are enrolled in the study. First, three typical convolutional neutral networks (CNNs) – VGG, ResNet and DenseNet – are trained to distinguish PsP and TTP on the dataset. Subsequently, we use the class-specific gradient information from convolutional layers to highlight the important regions in MRI. Radiological experts are then recruited to select the most lesion-relevant layer of each CNN. Finally, the selected layers are utilized to guide the construction of multi-scale ensemble CNN. The classified accuracy of the presented network is 90.20%, the promotion of specificity reaches more than 20%. The results demonstrate that network transparency and ensemble can enhance the reliability and accuracy of CNNs. The presented network is promising for the diagnosis of PsP and TTP.

📄 Content

Transparency guided ensemble convolutional neural networks for stratification of pseudoprogression and true progression of glioblastoma multiform

Xiaoming Liub, Michael D. Chand, Xiaobo Zhouc, Xiaohua Qiana

a Institute for Medical Imaging Technology, School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai 200030, China b College of electronic science and engineering, Jilin University c School of Biomedical Informatics, The University of Texas Health Science Center at Houston, Houston, TX 77030, USA d Department of Radiology, Wake Forest School of Medicine, Winston-Salem, NC 27157, USA

ABSTRACT Pseudoprogression (PsP) is an imitation of true tumor progression (TTP) in patients with glioblastoma multiform (GBM). Differentiating them is a challenging and time-consuming task for radiologists. Although deep neural networks can automatically diagnose PsP and TTP, interpretability shortage is always the Achilles’s heel. To overcome these shortcomings and win physician’s trust, we propose a transparency guided ensemble convolutional neural network to automatically stratify PsP and TTP on magnetic resonance imaging (MRI). A total of 84 patients with GBM are enrolled in the study. First, three typical convolutional neutral networks (CNNs)- VGG, ResNet and DenseNet- are trained to distinguish PsP and TTP on the dataset. Subsequently, we use the class-specific gradient information from convolutional layers to highlight the important regions in MRI. Radiological experts are then recruited to select the most lesion-relevant layer of each CNN. Finally, the selected layers are utilized to guide the construction of multi-scale ensemble CNN. The classified accuracy of the presented network is 90.20%, the promotion of specificity reaches more than 20%. The results demonstrate that network transparency and ensemble can enhance the reliability and accuracy of CNNs. The presented network is promising for the diagnosis of PsP and TTP.

1. Introduction Glioblastoma multiforme (GBM) is the most common malignant central nervous system tumor in adults.1 Currently, the standard treatment for GBM is surgical resection followed by radiation therapy with concurrent temozolomide (TMZ) and adjuvant TMZ. Despite of the benefits for patients, such treatment faces a new dilemma: the incidence of pseudoprogression (PsP). PsP is caused by post-treatment reactions, such as inflammation, ischemia or radiation necrosis 2 and will occur in about a third of all patients with GBM. PsP is a contrast enhancement that mimics early tumor progression, but unlike true tumor progression (TTP), it can improve or stabilize spontaneously without intervention. 3 Accurate diagnosis of PsP and TTP is critical because the results may directly influence the therapy strategy option and overall survival of patients. Although brain tumor biopsies can effectively differentiate PsP, the invasive procedure and second surgery may cause more risks of patients. Besides, follow-up imaging can provide an accurate identification, but it may discourage the patients of TTP to capture the best treatment opportunity. Diffusion tensor imaging (DTI) is a type of magnetic resonance imaging (MRI) which is good at anisotropy measurement. Since brain tissues with PsP have lower FA values than TTP, DTI is regarded as a potential way to truly reflect PsP and TTP.4,5,6 However, because PsP is very similar with TTP in not only intensity but shapes, it is still difficult for radiologists identifying them by DTI in the clinic. Therefore, a fast and non- invasive method is extremely needed to diagnose PsP and TTP. Techniques of computer-aided diagnosis (CAD) can fully take use of the gray information in medical images and then extract huge amount of features. The combination of minable features and advanced algorithms is more likely to outperform human in some medical diagnosis cases. Several attempts of CAD, such as parametric response maps 7, 8, 9, gray-level co-occurrence matrix 10 and dictionary learning 6are applied in the diagnosis of PsP and TTP. However, these methods cannot utilize high-dimensional features and thus fail to capture subtle differences between PsP and TTP in MRIs. Overall, the performances on diagnosis system of PsP and TTP are expected to be improved.
   Convolutional neutral networks (CNNs) are the emerging technique for image classification. CNNs are composed of many filter-based layers and can learn high-level features of the input images. CNNs consist of three types of layers: convolutional, pooling and fully connected. For each convolutional layer l, the input image is firstly convolved with several convolutional kernels W and then corresponding bias b is added. If the number of the kernel is K, the feature map 𝑋" of l can be computed as: 𝑋"

= 𝜎(𝑊"

#)* ∙ 𝑋#)* + 𝑏" #)*). σ(∙) is the activation function which performs non- linear transformation of 𝑋". Generally, convolutional layers are f

This content is AI-processed based on ArXiv data.