Complete removal of cancer tumors with a negative specimen margin during lumpectomy is essential to reduce breast cancer recurrence. However, 2D radiography, the current method used to assess intraoperative specimen margin status, has limited accuracy, resulting in nearly one in four patients needing repeat surgery. This study aims to develop a deep learning model that improves the detection of positive margins in intraoperative breast lumpectomy specimens on radiographs. We annotated the lumpectomy radiograph images with masking that denotes regions of known malignancy, non-malignant tissue, and the areas of pathology-confirmed positive margin. We propose a pretraining strategy, namely Forward-Forward Contrastive Learning (FFCL) with both local and global-level contrastive learning. Experimental results on our annotated breast radiographs demonstrate the effectiveness of our FFCL method in detecting positive margins from intraoperative radiographs of breast lumpectomy specimens.
CT diagnostic imaging is a major contributor to ionizing radiation exposure in the United States. Unfortunately, a reduction in radiation dose often results in degraded image quality. Automatic Exposure Control (AEC) is the most commonly used method to balance image quality and dose in x-ray CT, generally by modifying the scan’s tube current modulation (TCM) parameters. To allow current AEC techniques to be better personalized to the patient size, organ dose, and clinical task, our team previously proposed Scout-Dose and Scout-IQA to prospectively estimate dose and noise from frontal and lateral scouts, scan range, and TCM map. In this study, we evaluate for the first time the performance of our scout-based organ dose and noise predictions in an optimization framework to prospectively determine real-time, personalized TCM maps from a patient’s acquired scouts and scan ranges.
Automatic exposure control based on tube current modulation (TCM) can effectively reduce dose while maintaining image quality. Conventional TCM uses total exposure from the tube and noise in the center of CT slices as surrogates of dose and image quality, respectively. In this abstract, we present an automated method to optimize TCM at the organ level, offering increased flexibility and aligning with the concept of organ-specific radiation risk assessment. We applied our method to a retrospective CT dataset and incorporated automatic organ segmentation, Monte Carlo simulation for dose calculation, and an empirical model for noise estimation. This method was fully automated and readily scalable to massive clinical data, allowing the generation of ground-truth data for any data-driven approach to prospective planning, including methods utilizing scout images.
Automatic Target Recognition (ATR) is a valuable application of computer vision that traditionally requires copious and tedious labeling through supervised learning. This research explored if ATR can be performed on satellite imagery at a comparable accuracy to a fully supervised baseline model with a considerably smaller subset of data labelled, on the order of 10%, using a recently developed semi-supervised technique, contrastive learning. Supervised contrastive loss was explored and compared to traditional cross entropy loss. Supervised contrastive loss was found to perform significantly better with a subset of the data labelled on the XView dataset, a publicly available dataset of satellite imagery captured with .3 meter ground sampling. The caveats when nothing and everything is labelled were additionally explored.
Considering the potential radiation effect on patients in computed tomography (CT) imaging, it is desirable to reduce the radiation dose. Reduction in dose incurs degradation in image quality and possible reduced diagnostic performance. CT image quality needs to be maintained at standards sufficient for effective clinical reading. Therefore, the dose reduction should be guided by the potential image quality. Unfortunately, currently available CT image quality assessment (IQA) tools are based on maintaining a uniform image quality or use the CT exams themselves to retrospectively determine image quality. A robust and comprehensive IQA metric should represent the image quality of each patient at the organ level, and before the CT exams. Towards this objective, we devise a fully-automated, end-to-end deep learning-based solution to perform real-time, patient-specific, organ-level image quality prediction of CT scans. Leveraging the 2D scout (frontal and lateral) images of the actual patients, which are routinely acquired prior to the CT scan, our proposed Scout-IQA model estimates the patient-specific mean noise in real-time for six different organs. Our experimental evaluation on real patient data demonstrates the effectiveness of our Scout model not only in real-time noise estimation (only 6 ms on average per scan), but also as a potential tool for optimizing CT radiation dose in individual patients.
CT image quality is reliant on radiation dose, as low dose CT (LDCT) scans contain increased noise in images. This compromises the diagnostic performance on such scans. Therefore, it is desirable to perform Image Quality Assessment (IQA) prior to diagnostic use of CT scans. Often, image quality is assessed with full-reference methods, where a LDCT is algorithmically compared against its full dose counterpart. However due to health concerns, acquiring full dose CT scans is challenging and not desirable. As an alternative, non-reference IQA (NR-IQA) can be performed. Moreover, IQA at the pixel level is important, as most IQA methods only provide a global assessment, which means localized regions of interest cannot be specifically assessed. A solution for localized-IQA is to produce visually-interpretable quality maps. Deep learning methods could be employed by leveraging computer vision techniques, such as Self-Supervised learning (SSL). In this work, we propose Noise2Quality (N2Q)—a novel self-supervised, non-reference, pixel-wise image quality assessment model to predict IQA maps from LDCTs. Self-supervised dose level prediction as an auxiliary task further improves the model performance. Our experimental evaluation both qualitatively and quantitatively demonstrates the effectiveness of the model in accurately predicting IQA maps over various baselines.
KEYWORDS: Magnetic resonance imaging, 3D modeling, Epilepsy, Brain, Neuroimaging, Image quality, 3D image processing, Pathology, Data modeling, Medical imaging
During childhood, neurological involvement in tuberous sclerosis complex (TSC) is a leading cause of death. Neurological involvement, including epilepsy, can cause significant long-term sequelae in children. Brain involvement in TSC can be detected by magnetic resonance imaging (MRI). Still, neuroimaging analysis is time-and labor-intensive, begging the need for automated approaches to these tasks to improve speed, accuracy, and availability. We explored the general feasibility of using three-dimensional convolutional neural networks (CNNs) to automatically enhance image diagnosis quality and consistency to identify anatomical abnormalities in TSC children. We trained the 3D CNN on axial T1-weighted, axial T2-weighted FLAIR, and 3DT1-FSPGR weighted images from 296 TSC and 245 Normal cases from birth to 8 years of age acquired at LeBonheur Children’s Hospital. In the best performing approach, we achieved an accuracy of 0.86 [95% CI:0.76-0.97] with 0.95% AUC. The code can be found in https://github.com/shabanian2018/TSC3DCNN
Anthropomorphic software breast phantoms are generated by simulating breast anatomy. Virtual Clinical Trial (VCT) tools are developed for evaluating novel imaging modalities, based on anthropomorphic breast phantoms. Simulation of breast anatomical structures requires informed selection of parameters, which is crucial for the simulation realism. Our goal is to optimize the parameter selection based upon the analysis of clinical images.
Adipose compartments defined by Cooper’s ligaments significantly contribute to breast image texture (parenchymal pattern) which affects image interpretation and lesion detection. We have investigated the distribution and orientation of compartments segmented from CT images of a mastectomy specimen. Ellipsoidal fitting was applied to 205 segmented compartments, by matching the moments of inertia. The goodness-of-fit was measured by calculating Dice coefficients. Compartment size, shape, and orientation were characterized by estimating the volume, axis ratio, and Euler’s angles of fitted ellipsoids. Potential correlations between estimated parameters were tested.
We found that the adipose compartments are well approximated with ellipsoids (average Dice coefficient of 0.79). The compartment size is correlated with the barycenter-chest wall distance (r=0.235, p-value<0.001). The goodness-of-fit to ellipsoids is correlated to the compartment shape (r=0.344, p-value<0.001). The shape is also correlated with barycenter coordinates. The compartment orientation is correlated to their size (Euler angle α: r=0.188, p-value=0.007; angle β: r=0.156, p-value=0.025) and the barycenter-chest wall distance (r=0.159, p-value=0.023). These results from the characterization of adipose compartments and the observed correlations could help improve the realism of simulated breast anatomy.
Anthropomorphic breast phantoms are important tools for a wide range of tasks including pre-clinical validation of novel imaging techniques. In order to improve the realism in the phantoms, assessment of simulated anatomical structures is crucial. Thickness of simulated Cooper’s ligaments influences the percentage of dense tissue, as well as qualitative and quantitative properties of simulated images.
We introduce three methods (2-dimensional watershed, 3-dimensional watershed, and facet counting) to assess the thickness of the simulated Cooper’s ligaments in the breast phantoms. For the validation of simulated phantoms, the thickness of ligaments has been measured and compared with the input thickness values. These included a total of 64 phantoms with nominal ligament thicknesses of 200, 400, 600, and 800 μm.
The 2-dimensional and 3-dimensional watershed transformations were performed to obtain the median skeleton of the ligaments. In the 2-dimensional watershed, the median skeleton was found cross-section by cross-section, while the skeleton was found for the entire 3-dimensional space in the 3-dimensional watershed. The thickness was calculated by taking the ratio of the total volume of ligaments and the volume of median skeleton. In the facet counting method, the ligament thickness was estimated as a ratio between estimated ligaments’ volume and average ligaments’ surface area.
We demonstrated that the 2-dimensional watershed technique overestimates the ligament thickness. Good agreement was found between the facet counting technique and the 3-dimensional watershed for assessing thickness. The proposed techniques are applicable for ligaments’ thickness estimation on clinical breast images, provided segmentation of Cooper’s ligaments has been performed.
Virtual clinical trials (VCTs) were introduced as a preclinical alternative to clinical imaging trials, and for the evaluation of breast imaging systems. Realism in computer models of breast anatomy (software phantoms), critical for VCT performance, can be improved by optimizing simulation parameters based on the analysis of clinical images. We optimized the simulation to improve the realism of simulated tissue compartments, defined by the breast Cooper’s ligaments. We utilized the anonymized, previously acquired CT images of a mastectomy specimen to manually segment 205 adipose compartments. We generated 1,440 anthropomorphic breast phantoms based on octree recursive partitioning. These phantoms included variations of simulation parameters—voxel size, number of compartments, percentage of dense tissue, and shape and orientation of the compartments. We compared distributions of the compartment volumes in segmented CT images and phantoms using Kolmogrov-Smirnov (KS) distance, Kullback-Leibler (KL) divergence and a novel distance metric (based on weighted sum of distribution descriptors differences). We identified phantoms with the size distributions closest to CT images. For example, KS resulted in the phantom with 1000 compartments, ligament thickness of 0.4 mm and skin thickness of 12 mm. We applied multilevel analysis of variance (ANOVAN) to these distance measures to identify parameters that most significantly influence the simulated compartment size distribution. We have demonstrated an efficient method for the optimization of phantom parameters to achieve realistic distribution of adipose compartment size. The proposed methodology could be extended to other phantom parameters (e.g., ligaments and skin thicknesses), to further improve realism of the simulation and VCTs.
Anthropomorphic software breast phantoms have been utilized for preclinical quantitative validation of breast imaging
systems. Efficacy of the simulation-based validation depends on the realism of phantom images. Anatomical
measurements of the breast tissue, such as the size and distribution of adipose compartments or the thickness of Cooper’s
ligaments, are essential for the realistic simulation of breast anatomy. Such measurements are, however, not readily
available in the literature. In this study, we assessed the statistics of adipose compartments as visualized in CT images of
a total mastectomy specimen. The specimen was preserved in formalin, and imaged using a standard body CT protocol
and high X-ray dose. A human operator manually segmented adipose compartments in reconstructed CT images using
ITK-SNAP software, and calculated the volume of each compartment. In addition, the time needed for the manual
segmentation and the operator’s confidence were recorded. The average volume, standard deviation, and the probability
distribution of compartment volumes were estimated from 205 segmented adipose compartments. We also estimated the
potential correlation between the segmentation time, operator’s confidence, and compartment volume. The statistical
tests indicated that the estimated compartment volumes do not follow the normal distribution. The compartment volumes
are found to be correlated with the segmentation time; no significant correlation between the volume and the operator
confidence. The performed study is limited by the mastectomy specimen position. The analysis of compartment volumes
will better inform development of more realistic breast anatomy simulation.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.