Pulmonary Nodule Classification Using a Multiview Residual Selective Kernel Network.

Lung cancer is one of the leading causes of death worldwide and early detection is crucial to reduce the mortality. A reliable computer-aided diagnosis (CAD) system can help facilitate early detection of malignant nodules. Although existing methods provide adequate classification accuracy, there is still room for further improvement. This study is dedicated to investigating a new CAD scheme for predicting the malignant likelihood of lung nodules in computed tomography (CT) images in light of a deep learning strategy. Conceived from the residual learning and selective kernel, we investigated an efficient residual selective kernel (RSK) block to handle the diversity of lung nodules with various shapes and obscure structures. Founded on this RSK block, we established a multiview RSK network (MRSKNet), to which three anatomical planes in the axial, coronal, and sagittal directions were fed. To reinforce the classification efficiency, seven handcrafted texture features with a filter-like computation strategy were explored, among which the homogeneity (HOM) feature maps are combined with the corresponding intensity CT images for concatenation input, leading to an improved network architecture. Evaluated on the public benchmark Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) challenge database with ten-fold cross validation of binary classification, our experimental results indicated high area under receiver operating characteristic (AUC) and accuracy scores. A better compromise between recall and specificity was struck using the suggested concatenation strategy comparing to many state-of-the-art approaches. The proposed pulmonary nodule classification framework exhibited great efficacy and achieved a higher AUC of 0.9711. The association of handcrafted texture features with deep learning models is promising in advancing the classification performance. The developed pulmonary nodule CAD network architecture is of potential in facilitating the diagnosis of lung cancer for further image processing applications.

Automatic segmentation of high-risk clinical target volume for tandem-and-ovoids brachytherapy patients using an asymmetric dual-path convolutional neural network.

PURPOSES: Preimplant diagnostic magnetic resonance imaging is the gold standard for image-guided tandem-and-ovoids (T&O) brachytherapy for cervical cancer. However, high dose rate brachytherapy planning is typically done on postimplant CT-based high-risk clinical target volume (HR-CTVCT ) because the transfer of preimplant Magnetic resonance (MR)-based HR-CTV (HR-CTVMR ) to the postimplant planning CT is difficult due to anatomical changes caused by applicator insertion, vaginal packing, and the filling status of the bladder and rectum. This study aims to train a dual-path convolutional neural network (CNN) for automatic segmentation of HR-CTVCT on postimplant planning CT with guidance from preimplant diagnostic MR. METHODS: Preimplant T2-weighted MR and postimplant CT images for 65 (48 for training, eight for validation, and nine for testing) patients were retrospectively solicited from our institutional database. MR was aligned to the corresponding CT using rigid registration. HR-CTVCT and HR-CTVMR were manually contoured on CT and MR by an experienced radiation oncologist. All images were then resampled to a spatial resolution of 0.5 × 0.5 × 1.25 mm. A dual-path 3D asymmetric CNN architecture with two encoding paths was built to extract CT and MR image features. The MR was masked by HR-CTVMR contour while the entire CT volume was included. The network put an asymmetric weighting of 18:6 for CT: MR. Voxel-based dice similarity coefficient (DSCV ), sensitivity, precision, and 95% Hausdorff distance (95-HD) were used to evaluate model performance. Cross-validation was performed to assess model stability. The study cohort was divided into a small tumor group (<20 cc), medium tumor group (20-40 cc), and large tumor group (>40 cc) based on the HR-CTVCT for model evaluation. Single-path CNN models were trained with the same parameters as those in dual-path models. RESULTS: For this patient cohort, the dual-path CNN model improved each of our objective findings, including DSCV , sensitivity, and precision, with an average improvement of 8%, 7%, and 12%, respectively. The 95-HD was improved by an average of 1.65 mm compared to the single-path model with only CT images as input. In addition, the area under the curve for different networks was 0.86 (dual-path with CT and MR) and 0.80 (single-path with CT), respectively. The dual-path CNN model with asymmetric weighting achieved the best performance with DSCV of 0.65 ± 0.03 (0.61-0.70), 0.79 ± 0.02 (0.74-0.85), and 0.75 ± 0.04 (0.68-0.79) for small, medium, and large group. 95-HD were 7.34 (5.35-10.45) mm, 5.48 (3.21-8.43) mm, and 6.21 (5.34-9.32) mm for the three size groups, respectively. CONCLUSIONS: An asymmetric CNN model with two encoding paths from preimplant MR (masked by HR-CTVMR ) and postimplant CT images was successfully developed for automatic segmentation of HR-CTVCT for T&O brachytherapy patients.

Quantitative Characterization of Tumor Proximity to Stem Cell Niches: Implications on Recurrence and Survival in GBM Patients.

PURPOSE: Emerging evidence has linked glioblastoma multiforme (GBM) recurrence and survival to stem cell niches (SCNs). However, the traditional tumor-ventricle distance is insufficiently powered for an accurate prediction. We aimed to use a novel inverse distance map for improved prediction. METHODS AND MATERIALS: Two T1-magnetic resonance imaging data sets were included for a total of 237 preoperative scans for prognostic stratification and 55 follow-up scans for recurrent pattern identification. SCN, including the subventricular zone (SVZ) and subgranular zone (SGZ), were manually defined on a standard template. A proximity map was generated using the summed inverse distances to all SCN voxels. The mean and maximum proximity scores (PSm-SCN and PSmax-SCN) were calculated for each primary/recurrent tumor, deformably transformed into the template. The prognostic capacity of proximity score (PS)-derived metrics was assessed using Cox regression and log-rank tests. To evaluate the impact of SCNs on recurrence patterns, we performed group comparisons of PS-derived metrics between the primary and recurrent tumors. For comparison, the same analyses were conducted on PS derived from SVZ alone and traditional edge/center-to-ventricle metrics. RESULTS: Among all SCN-derived features, PSm-SCN was the strongest survival predictor (P < .0001). PSmax-SCN was the best in risk stratification, using either evenly sorted (P = .0001) or k-means clustering methods (P = .0045). PS metrics based on SVZ only also correlated with overall survival and risk stratification, but to a lesser degree of significance. In contrast, edge/center-to-ventricle metrics showed weak to no prediction capacities in either task. Moreover, PSm-SCN,PSm-SVZ, and center-to-ventricle metrics revealed a significantly closer SCN distribution of recurrence than primary tumors. CONCLUSIONS: We introduced a novel inverse distance-based metric to comprehensively capture the anatomic relationship between GBM tumors and SCN zones. The derived metrics outperformed traditional edge or center distance-based measurements in overall survival prediction, risk stratification, and recurrent pattern differentiation. Our results reveal the potential role of SGZ in recurrence aside from SVZ.

Brain MR Image Restoration Using an Automatic Trilateral Filter With GPU-Based Acceleration.

OBJECTIVE: Noise reduction in brain magnetic resonance (MR) images has been a challenging and demanding task. This study develops a new trilateral filter that aims to achieve robust and efficient image restoration. METHODS: Extended from the bilateral filter, the proposed algorithm contains one additional intensity similarity funct-ion, which compensates for the unique characteristics of noise in brain MR images. An entropy function adaptive to intensity variations is introduced to regulate the contributions of the weighting components. To hasten the computation, parallel computing based on the graphics processing unit (GPU) strategy is explored with emphasis on memory allocations and thread distributions. To automate the filtration, image texture feature analysis associated with machine learning is investigated. Among the 98 candidate features, the sequential forward floating selection scheme is employed to acquire the optimal texture features for regularization. Subsequently, a two-stage classifier that consists of support vector machines and artificial neural networks is established to predict the filter parameters for automation. RESULTS: A speedup gain of 757 was reached to process an entire MR image volume of 256 × 256 × 256 pixels, which completed within 0.5 s. Automatic restoration results revealed high accuracy with an ensemble average relative error of 0.53 ± 0.85% in terms of the peak signal-to-noise ratio. CONCLUSION: This self-regulating trilateral filter outperformed many state-of-the-art noise reduction methods both qualitatively and quantitatively. SIGNIFICANCE: We believe that this new image restoration algorithm is of potential in many brain MR image processing applications that require expedition and automation.