Quantitative and Visual Analysis of Data Augmentation and Hyperparameter Optimization in Deep Learning-Based Segmentation of Low-Grade Glioma Tumors Using Grad-CAM.

This study executes a quantitative and visual investigation on the effectiveness of data augmentation and hyperparameter optimization on the accuracy of deep learning-based segmentation of LGG tumors. The study employed the MobileNetV2 and ResNet backbones with atrous convolution in DeepLabV3+ structure. The Grad-CAM tool was also used to interpret the effect of augmentation and network optimization on segmentation performance. A wide investigation was performed to optimize the network hyperparameters. In addition, the study examined 35 different models to evaluate different data augmentation techniques. The results of the study indicated that incorporating data augmentation techniques and optimization can improve the performance of segmenting brain LGG tumors up to 10%. Our extensive investigation of the data augmentation techniques indicated that enlargement of data from 90° and 225° rotated data,up to down and left to right flipping are the most effective techniques. MobilenetV2 as the backbone,"Focal Loss" as the loss function and "Adam" as the optimizer showed the superior results. The optimal model (DLG-Net) achieved an overall accuracy of 96.1% with a loss value of 0.006. Specifically, the segmentation performance for Whole Tumor (WT), Tumor Core (TC), and Enhanced Tumor (ET) reached a Dice Similarity Coefficient (DSC) of 89.4%, 70.1%, and 49.9%, respectively. Simultaneous visual and quantitative assessment of data augmentation and network optimization can lead to an optimal model with a reasonable performance in segmenting the LGG tumors.

Automatic segmentation of glioblastoma multiform brain tumor in MRI images: Using Deeplabv3+ with pre-trained Resnet18 weights.

PURPOSE: To assess the effectiveness of deep learning algorithms in automated segmentation of magnetic resonance brain images for determining the enhanced tumor, the peri-tumoral edema, the necrotic/ non-enhancing tumor, and Normal tissue volumes. METHODS AND MATERIALS: A new deep neural network algorithm, Deep-Net, was developed for semantic segmentation of the glioblastoma tumors in MR images, using the Deeplabv3+ architecture, and the pre-trained Resnet18 initial weights. The MR image Dataset used for training the network was taken from the BraTS 2020 training set, with the ground truth labels for different tumor subregions manually drawn by a group of expert neuroradiologists. In this work, two multi-modal MRI scans, i.e., T1ce and FLAIR of 293 patients with high-grade glioma (HGG), were used for deep network training (Deep-Net). The performance of the network was assessed for different hyper-parameters, to obtain the optimum set of parameters. The similarity scores were used for the evaluation of the optimized network. RESULTS: According to the results of this study, epoch #37 is the optimum epoch giving the best global accuracy (97.53%), and loss function (0.14). The Deep-Net sensitivity in the delineation of the enhanced tumor is more than 90%. CONCLUSIONS: The results indicate that the Deep-Net was able to segment GBM tumors with high accuracy.

DCE-MRI prediction of survival time for patients with glioblastoma multiforme: using an adaptive neuro-fuzzy-based model and nested model selection technique.

This pilot study investigates the construction of an Adaptive Neuro-Fuzzy Inference System (ANFIS) for the prediction of the survival time of patients with glioblastoma multiforme (GBM). ANFIS is trained by the pharmacokinetic (PK) parameters estimated by the model selection (MS) technique in dynamic contrast enhanced-magnetic resonance imaging (DCE-MRI) data analysis, and patient age. DCE-MRI investigations of 33 treatment-naïve patients with GBM were studied. Using the modified Tofts model and MS technique, the following physiologically nested models were constructed: Model 1, no vascular leakage (normal tissue); Model 2, leakage without efflux; Model 3, leakage with bidirectional exchange (influx and efflux). For each patient, the PK parameters of the three models were estimated as follows: blood plasma volume (vp ) for Model 1; vp and volume transfer constant (Ktrans ) for Model 2; vp , Ktrans and rate constant (kep ) for Model 3. Using Cox regression analysis, the best combination of the estimated PK parameters, together with patient age, was identified for the design and training of ANFIS. A K-fold cross-validation (K = 33) technique was employed for training, testing and optimization of ANFIS. Given the survival time distribution, three classes of survival were determined and a confusion matrix for the correct classification fraction (CCF) of the trained ANFIS was estimated as an accuracy index of ANFIS's performance. Patient age, kep and ve (Ktrans /kep ) of Model 3, and Ktrans of Model 2, were found to be the most effective parameters for training ANFIS. The CCF of the trained ANFIS was 84.8%. High diagonal elements of the confusion matrix (81.8%, 90.1% and 81.8% for Class 1, Class 2 and Class 3, respectively), with low off-diagonal elements, strongly confirmed the robustness and high performance of the trained ANFIS for predicting the three survival classes. This study confirms that DCE-MRI PK analysis, combined with the MS technique and ANFIS, allows the construction of a DCE-MRI-based fuzzy integrated predictor for the prediction of the survival of patients with GBM.

An adaptive model for rapid and direct estimation of extravascular extracellular space in dynamic contrast enhanced MRI studies.

Extravascular extracellular space (ve ) is a key parameter to characterize the tissue of cerebral tumors. This study introduces an artificial neural network (ANN) as a fast, direct, and accurate estimator of ve from a time trace of the longitudinal relaxation rate, ΔR1 (R1  = 1/T1 ), in DCE-MRI studies. Using the extended Tofts equation, a set of ΔR1 profiles was simulated in the presence of eight different signal to noise ratios. A set of gain- and noise-insensitive features was generated from the simulated ΔR1 profiles and used as the ANN training set. A K-fold cross-validation method was employed for training, testing, and optimization of the ANN. The performance of the optimal ANN (12:6:1, 12 features as input vector, six neurons in hidden layer, and one output) in estimating ve at a resolution of 10% (error of ±5%) was 82%. The ANN was applied on DCE-MRI data of 26 glioblastoma patients to estimate ve in tumor regions. Its results were compared with the maximum likelihood estimation (MLE) of ve . The two techniques showed a strong agreement (r = 0.82, p < 0.0001). Results implied that the perfected ANN was less sensitive to noise and outperformed the MLE method in estimation of ve .