Joseph Lo

© 2019, Springer Nature Switzerland AG. Generative Adversarial Networks (GANs) have found applications in natural image synthesis and begin to show promises generating synthetic medical images. In many cases, the ability to perform controlled image synthesis using masked priors such as shape and size of organs is desired. However, mask-guided image synthesis is challenging due to the pixel level mask constraint. While the few existing mask-guided image generation approaches suffer from the lack of fine-grained texture details, we tackle the issue of mask-guided stochastic image synthesis via mask embedding. Our novel architecture first encodes the input mask as an embedding vector and then inject these embedding into the random latent vector input. The intuition is to classify semantic masks into partitions before feature up-sampling for improved sample space mapping stability. We validate our approach on a large dataset containing 39,778 patients with 443,556 negative screening Full Field Digital Mammography (FFDM) images. Experimental results show that our approach can generate realistic high-resolution (256 × 512 ) images with pixel-level mask constraints, and outperform other state-of-the-art approaches.

OBJECTIVE: The goal of this study is to use adjunctive classes to improve a predictive model whose performance is limited by the common problems of small numbers of primary cases, high feature dimensionality, and poor class separability. Specifically, our clinical task is to use mammographic features to predict whether ductal carcinoma in situ (DCIS) identified at needle core biopsy will be later upstaged or shown to contain invasive breast cancer. METHODS: To improve the prediction of pure DCIS (negative) versus upstaged DCIS (positive) cases, this study considers the adjunctive roles of two related classes: atypical ductal hyperplasia (ADH), a non-cancer type of breast abnormity, and invasive ductal carcinoma (IDC), with 113 computer vision based mammographic features extracted from each case. To improve the baseline Model A classification of pure vs. upstaged DCIS, we designed three different strategies (Models B, C, D) with different ways of embedding features or inputs. RESULTS: Based on ROC analysis, the baseline Model A performed with AUC of 0.614 (95% CI, 0.496-0.733). All three new models performed better than the baseline, with domain adaptation (Model D) performing the best with an AUC of 0.697 (95% CI, 0.595-0.797). CONCLUSION: We improved the prediction performance of DCIS upstaging by embedding two related pathology classes in different training phases. SIGNIFICANCE: The three new strategies of embedding related class data all outperformed the baseline model, thus demonstrating not only feature similarities among these different classes, but also the potential for improving classification by using other related classes.

© SPIE. Downloading of the abstract is permitted for personal use only. Custom 3D printed physical phantoms are desired for testing the limits of medical imaging, and for providing patientspecific information. This work focuses on the development of low-cost, open source fused filament fabrication for printing of physical phantoms with the structure and contrast of human anatomy in computed tomography (CT) images. Specifically, this paper introduces the concept of using a porous 3D printed layer as a background into which additional material can be printed to control the position-dependent contrast. By using this method, eight levels of contrast were printed with a single material.

© SPIE. Downloading of the abstract is permitted for personal use only. The purpose of this study was to develop a robust, automated multi-organ segmentation model for clinical adult and pediatric CT and implement the model as part of a patient-specific safety and quality monitoring system. 3D convolutional neural network (Unet) models were setup to segment 30 different organs and structures at the diagnostic image resolution. For each organ, 200 manually-labeled cases were used to train the network, fitting it to different clinical imaging resolutions and contrast enhancement stages. The dataset was randomly shuffled, and divided with 6/2/2 train/validation/test set split. The model was deployed to automatically segment 1200 clinical CT images as a demonstration of the utility of the method. Each case was made into a patient-specific phantom based on the segmentation masks, with unsegmented organs and structures filled in by deforming a template XCAT phantom of similar anatomy. The organ doses were then estimated using a validated scanner-specific MC-GPU package using the actual scan information. The segmented organ information was likewise used to assess contrast, noise, and detectability index within each organ. The neural network segmentation model showed dice similarity coefficients (DSC) above 0.85 for the majority of organs. Notably, the lungs and liver showed a DSC of 0.95 and 0.94, respectively. The segmentation results produced patient-specific dose and quality values across the tested 1200 patients with representative the histogram distributions. The measurements were compared in global-to-organ (e.g. CTDvol vs. liver dose) and organ-to-organ (e.g. liver dose vs. spleen dose) manner. The global-to-organ measurements (liver dose vs. CTDIvol: o-'. = 0.62; liver vs. global d': o'. = 0.78; liver vs. global noise: o'. = 0.55) were less correlated compared to the organ-to-organ measurements (liver vs. spleen dose: o'. = 0.93; liver vs. spleen d': o'. = 0.82; liver vs. spleen noise: o'. = 0.78). This variation of measurement is more prominent for iterative reconstruction kernel compared to the filtered back projection kernel (liver vs. global noise: o'.o4o'. = 0.47 vs. o'.oouo F = 0.75; liver vs. global d': o'.o1/4o'. = 0.74 vs. o'.oouo'F = 0.83). The results can help derive meaningful relationships between image quality, organ doses, and patient attributes.