Browsing by Subject "CNN"

PurposeThe aim of this study was to address the limitations observed in past research, particularly the limited accuracy of individual deep learning or radiomics methods in small datasets. By developing a fusion approach that integrates the two techniques, we hypothesized that the performance in CT-based thyroid cancer diagnosis could be improved. Materials and Methods Eighty-five patients with thyroid tumors (58 malignant, 27 benign) who underwent CT scans were included in this study. The dataset was divided into training (70%) and testing (30%). A shallow CNN model, including five convolutional layers and two fully connected layers, was developed for tumor classification. Radiomics features were extracted and selected using the pyradiomics package and statistical tests (T-test, etc.). These features were then utilized to develop a Multiple Logistic Regression (MLR) model for tumor classification. The CNN and MLR models were combined using a fusion method that calculates the weighted sum of each diagnostic output for classification. The accuracy of the diagnostic methods was evaluated for both the individual and combined fusion models. The statistical significance of the weighted combination model was examined using the Wilcoxon-Test. Results The CNN model achieved an accuracy of 82.713%, and the MLR model achieved an accuracy of 76.596%. The accuracy of the fusion model reached 85.372%, suggested the improvement of performance of the fusion approach over the individual models. The Wilcoxon-Test yielded a W-Statistic of 19410.0 and a p-value of 〖2.96×10〗^(-14), which is below the threshold of 0.05. Conclusion A fusion model combining deep learning and radiomics methods was developed and showed improved accuracy in thyroid tumor diagnosis in a small dataset. The results showed a statistically significant difference between the fusion model and the individual models.

Purpose: Imaging data scarcity is an intrinsic challenge in medical research partly due to resource limitations and privacy concerns. Digital human phantom is one approach to mitigate such challenge. The aim of this study is to propose an innovative method combining Generative Adversarial Network (GAN) and Convolutional Neural Network (CNN) to virtually synthesize large scale and high-quality brain cortex digital phantoms.Method: High resolution brain MRIs (MP-RAGE) of 369 patients were used to generate brain segmentations using FreeSurfer software. Total of 14,200 segmented 2D cortex phantoms were extracted and served as ground truth. A GAN was developed to learn and synthesize new cortex phantoms. 28,955 phantoms were generated and manually classified into ‘real’ and ‘fake’ groups, where 14,364 were labeled as ‘real’ while rest were labeled as ‘fake’ due to mal/poor morphology and/or noise. To facilitate automatic phantom quality classification, a CNN was designed and trained on the total of 43,155 phantoms, where the training and testing split was 7:3. Receiver Operating Characteristic Curve (ROC) analysis and Area Under the Curve (AUC) evaluation were performed on the CNN. The GAN and CNN together constitute the automatic cortex digital phantom generation network. To visually evaluate the realism of the model-generated phantoms, two experiments were conducted: the first involved binary classification of real and generated phantoms, and the second added a ‘Cannot Decide’ option to the previous task. Eight volunteers were recruited for these experiments. In addition to visual assessments, the Fréchet Inception Distance (FID) was employed as a quantitative metric to evaluate the generated phantoms. Results: The CNN-based filtering network attained an accuracy of 97% and an AUC of 0.99. This integrated GAN generation and CNN classification network generated over 130,000 high-quality phantoms in less than one hour on a PC (NVIDIA GeForce GTX 1060), with only 6 phantoms contained minor noise on the phantom border. The morphological quality of the final phantoms received positive approval from visual evaluation experiments, demonstrating that the ability to distinguish synthetic from real phantoms in paired comparisons was about 50%. The FID score of 9.61 suggests a high level of diversity in the synthetic phantoms. Conclusion: This study demonstrated a novel and efficient method for generating a large-scale synthetic medical digital phantom by integrating two deep-learning models, GAN and CNN. The synthesized digital phantoms not only showed realistic morphology but also maintains commendable data diversity. Additionally, this approach offers a solution to the time-consuming and subjective nature of manually assessing the phantom quality generated by GAN networks. The digital phantoms synthesized have potential to be utilized in various phantom research areas.

The resurgence of Convolutional Neural Networks (CNNs) from the early foundational work is largely attributed to the advent of extensive manually labeled datasets, which has made it possible to train high-capacity models with strong generalization capabilities. However, the annotation cost for these datasets is often prohibitive, and so training CNNs on limited data in a fully-supervised setting remains a crucial problem. Data augmentation is a promising direction for improving generalization in scarce data settings.

We study foundational augmentation techniques, including Mixed Sample Data Augmentations (MSDAs) and a no-parameter variant of RandAugment termed Preset-RandAugment, in the fully supervised scenario. We observe that Preset-RandAugment excels in limited-data contexts while MSDAs are moderately effective. In order to explain this behaviour, we refine ideas about diversity and realism from prior work and propose new ways to measure them. We postulate an additional property when data is limited: augmentations should encourage faster convergence by helping the model learn stable and invariant low-level features, focusing on less class-specific patterns. We explain the effectiveness of Preset-RandAugment in terms of these properties and identify low-level feature transforms as a key contributor to performance.

Building on these insights, we introduce a novel augmentation technique called RandMSAugment that integrates complementary strengths of existing methods. It combines low-level feature transforms from Preset-RandAugment with interpolation and cut-and-paste from MSDA. We improve image diversity through added stochasticity in the mixing process. RandMSAugment significantly outperforms the competition on CIFAR-100, STL-10, and Tiny-Imagenet. With very small training sets (4, 25, 100 samples/class), RandMSAugment achieves compelling performance gains between 4.1\% and 6.75\%. Even with more training data (500 samples/class) we improve performance by 1.03\% to 2.47\%. We also incorporate RandMSAugment augmentations into a semi-supervised learning (SSL) framework and show promising improvements over the state-of-the-art SSL method, FlexMatch. The improvements are more significant when the number of labeled samples is smaller. RandMSAugment does not require hyperparameter tuning, extra validation data, or cumbersome optimizations.

Finally, we combine RandMSAugment with another powerful generalization tool, ensembling, for fully-supervised training with limited samples. We show additonal improvements on the 3 classification benchmarks, which range between 2\% and 5\%. We empirically demonstrate that the gains due to ensembling are larger when the individual networks have moderate accuracies \ie outside of the low and high extremes.Furthermore, we introduce a simulation tool capable of providing insights about the maximum accuracy achievable through ensembling, under various conditions.

Objective

Approach

Main results

Significance

Purpose: The purpose is to investigate the feasibility of using Convolutional Neural Network (CNN) to register phase-to-phase deformation vector field (DVF) of lung 4D Computed Tomography (CT) / Cone-Beam Computed Tomography (CBCT).

Methods: A Convolutional Neural Network (CNN) based deep learning method was built to directly register the deformation vector field from the individual phases images of patient 4D CT or 4D CBCT for 4D contouring, dose accumulation or target verification. . The input consists of image pairs while the output is the corresponding DVF that registers the image pairs. The network consisted of four convolutional layers, two average pooling layers and two fully connected layers. The loss function was half mean squared error. The centers of patch pairs are uniformly chosen across the lung and the number of samples were chosen to cover the majority movement of deformable vectors. The method was evaluated by using4-dimensional image volumes from 9 patients with lung cancer, such as 4D-CT, simulated 4D-CBCT reconstructed from DRR, and real 4D-CBCT reconstructed from real projections. In intra-patient study, These image volumes were sortedthe image volumes were sorted into different combinations, (1) training and testing samples from the same 4D-CT image volume, (2) training and testing samples from two 4D-CT volumes, (3) training and testing samples from 4D-CBCT volumes simulated by DRR from 4D-CT volumes, (4) training from 4D-CT and testing from 4D-CBCT reconstructed from primary projections, and (5) training and testing samples from two 4D-CBCT volumes reconstructed from primary projections. In inter-patient study, five 4D-CT volumes from five patients were used as the training set and the sixth patient 4D-CT volume was the testing set. The functionality of a well-trained network adapting new patient’s anatomy was tested. The coefficient of correlations between the prediction DVF and the ground truth DVF and the cross correlations between the target image and the ground truth deformed image, and the target image and the predicted deformed image were calculated. The registered images from predicted DVF and ground truth DVF were reconstructed and compared. One set being the training set and the other the testing set. The centers of patches (i.e. control points) are uniformly chosen across the lung. The limit of physical memory compromises the number of samples and size of patches.

Result: The ratio of cross correlation between predicted deformed image and target image to cross correlation between ground truth image and target image is 0.78 by averaging the intra-patient studies. For inter-patient study, this number is 0.62. In comparing the predicted deformed image and the ground truth image, major features such as diaphragms, lumen and main vessels are matched with each other.

Conclusion: CNN based regression model successfully learn the DVF from one image set, and the trained model can be successfully transferred to another data set, provided the high image quality in training sets and similar anatomic structure between both image sets.