Abstract PD11-02: Subtyping invasive carcinomas and high-risk lesions for machine learning based breast pathology

Background. The female mammary gland can develop a myriad of epithelial proliferative lesions including, high risk lesions, in-situ and invasive carcinomas. Identification of these pre-neoplastic and neoplastic conditions in biopsy specimens is crucial for proper patient management and may sometimes pose diagnostic challenges for pathologists. Recent research has shown that machine learning algorithms applied to whole slide images (WSI) can accurately detect and grade various cancers; herein, we devise and test a system that classifies the most common preneoplastic and neoplastic conditions of the female breast from WSIs. Design. De-identified slides were scanned on Leica AT2 whole slide scanners (20x; 0.5 µm/pixel) from MSK database were retrieved. Clinical diagnostic metadata were extracted from the pathology reports. Using a multi-label multiple-instance learning (ML-MIL) approach, a SE-ResNet50 Convolutional Neural Network (CNN) was trained to classify atypical lobular hyperplasia (ALH), atypical ductal hyperplasia (ADH), lobular carcinoma in situ (LCIS), ductal carcinoma in situ (DCIS), invasive lobular carcinoma (ILC), invasive ductal carcinoma (IDC). In additional morphological subtypes including apocrine, mucinous, solid papillary, micropapillary, and tubular carcinoma were trained. The system uses the WSI as an input and outputs a slide level class and heatmap for the presence of the trained classes. A validation dataset separate from the training set was used to assess performance of the trained model. Results. The CNN was trained on 9,751 surgical specimens (biopsy, 6,289; excision, 3,462) comprising 40,637 slides. The system was validated on 3,183 breast specimens (biopsy, 1,934; excision, 1,249) comprising 11,447 digital slides that were not included in the training of the CNN model. Validation performance in terms of Area Under Receiver Operating Characteristic Curve (AUROC) for each class is shown in Table 1. Conclusion. The trained CNN had a high performance in identifying the presence of ADH, ALH, DCIS, IDC, ILC, LCIS, and, apocrine, micropapillary, mucinous, solid papillary, and tubular carcinomas. Further studies expanding classes to include all clinically relevant lesions and morphologies are underway. In addition, the same approach can be used to detect microinvasions and calcifications in breast tissue.

Table 1.Area Under Receiver Operating Characteristic Curve for Breast Lesion ClassesClassNum. Positive (specimens)Num. Negative (specimens)AUROCADH24718640.903ALH20918860.950LCIS17520080.958DCIS81920920.956IDC52126620.956ILC7131120.934Apocrine2431590.931Micropapillary17230110.927Mucinous1231710.994Solid Papillary1531680.908Tubular carcinoma831750.990

Citation Format: Matthew G Hanna, Patricia Raciti, Alican Bozkurt, Ran Godrich, Julian Viret, Donghun Lee, Philippe Mathieu, Matthew Lee, Eugene Vorontsov, Tomer Sabo, Felipe C Geyer, Jorge S Reis-Filho, Leo Grady, Thomas Fuchs, Christopher Kanan. Subtyping invasive carcinomas and high-risk lesions for machine learning based breast pathology [abstract]. In: Proceedings of the 2021 San Antonio Breast Cancer Symposium; 2021 Dec 7-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2022;82(4 Suppl):Abstract nr PD11-02.

Abstract PD11-01: An artificial intelligence-based predictor of CDH1 biallelic mutations and invasive lobular carcinoma

Introduction: Invasive lobular carcinoma (ILC) is the most frequent special histologic subtype of breast cancer (BC). ILC is identifiable by pathologic assessment given its distinctive discohesive growth pattern, largely caused by CDH1 inactivation. Compared to common forms of BC, ILCs display lower responses to chemotherapy and selective estrogen receptor modulators. The low interobserver agreement for the diagnosis of ILC, however, renders the inclusion of histologic subtyping in therapeutic decision-making challenging. Artificial intelligence (AI)-based algorithms hold promise for improving pathologic diagnosis; their performance, however, depends on the ground truth labeling used. Here, we seek to develop an AI-based methodology for detection of ILC using ‘CDH1 biallelic mutations’ (i.e., mutation + loss-of-heterozygosity of the wild-type allele or two pathogenic somatic mutations) as ground truth, reasoning that in BC, >95% of CDH1 bi-allelic inactivation is found in ILCs.Materials and methods: We developed a convolutional neural network system to detect CDH1 biallelic genetic inactivation (AI-CDH1) using whole slide images (WSI) of 1,100 primary BCs with available targeted sequencing data. The model was trained using a 10-fold cross-validation method to detect biallelic mutations. The mean number of positive and negative samples in the training set was 85.2 (SD=2.57) and 562.8 (SD=10.51) per fold, respectively. The evaluation set consisted of a mean of 14.2 (SD=2.04) positive and 93.8 (SD=9.13) negative samples. We evaluated the performance of the AI-CDH1 classifier to predict the lobular phenotype and CDH1 status using original and revised labels, following a histopathologic re-review of the histologic type and CDH1 status curation. The latter was conducted by incorporating information on biallelic CDH1 inactivation beyond CDH1 mutations (homozygous deletions, deleterious structural rearrangements, and loss-of-heterozygosity and gene promoter methylation).Results: The AI-CDH1 classifier predicted biallelic CDH1 mutations with an area under the curve (AUC)=0.944 (95 CI: 0.925-0.963), sensitivity=91.6% and specificity=85.9%, PPV=49.8%, NPV=98.5% and accuracy=86.7%, and the original ‘lobular phenotype’ with an AUC=0.941 (95 CI: 0.922-0.960), sensitivity=89%, specificity=86.7%, PPV=55.6%, NPV=97.7% and accuracy=87.1%. Review of the CDH1 gene status revealed that 7/957 BCs lacking CDH1 biallelic mutations harbored biallelic CDH1 inactivation by promoter methylation, homozygous deletions or structural rearrangements. The AI-CDH1 classifier detected all seven reclassified BCs and predicted the revised CDH1 biallelic inactivation with an AUC=0.948 (95 CI: 0.930-0.966), sensitivity=92%, specificity=86.5%, PPV=52.3%, NPV=98.5% and accuracy=87.2%. Upon histologic re-review, which resulted in reclassification of 36/927 non-lobular BCs as ‘lobular’ and 5/173 ‘lobular’ BCs as ‘non-lobular’, the AI-CDH1 classifier detected the ‘lobular phenotype’ with an AUC=0.953 (95 CI: 0.935-0.971), sensitivity=90.7%, specificity=89.7%, PPV=66.8%, NPV=97.7% and accuracy=89.9%. Using the revised histologic re-classification and CDH1 biallelic inactivation status labels, the AI-CDH1 classifier predicted the lobular phenotype irrespective of CDH1 status (P>0.05).Conclusions: By training a machine learning system to detect ‘CDH1 biallelic mutations’, as ground truth rather than histologic diagnosis of lobular carcinoma, which might be confounded by human subjectivity, we developed an AI-based system that can detect ILCs accurately, providing a new paradigm for the development of AI-based cancer classification systems.

Citation Format: Jorge S Reis-Filho, Fresia Pareja, Fatemeh Derakhshan, David N Brown, Jillian Sue, Pier Selenica, Yi Kan Wang, Arnaud Da Cruz Paula, Monami Banerjee, Zahra Ebrahimzadeh, Manuel Isava, Matthew Lee, Ran Godrich, Adam Casson, Ruben Padron, George Shaikovski, Alexander van Eck, Antonio Marra, Higinio Dopeso, Hannah Y Wen, Edi Brogi, Matthew G Hanna, Chris Kanan, Jeremy D Kunz, Felipe C Geyer, Carla Leibowitz, David Klimstra, Leo Grady, Thomas J Fuchs. An artificial intelligence-based predictor of CDH1 biallelic mutations and invasive lobular carcinoma [abstract]. In: Proceedings of the 2021 San Antonio Breast Cancer Symposium; 2021 Dec 7-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2022;82(4 Suppl):Abstract nr PD11-01.

Independent real-world application of a clinical-grade automated prostate cancer detection system

Artificial intelligence (AI)‐based systems applied to histopathology whole‐slide images have the potential to improve patient care through mitigation of challenges posed by diagnostic variability, histopathology caseload, and shortage of pathologists. We sought to define the performance of an AI‐based automated prostate cancer detection system, Paige Prostate, when applied to independent real‐world data. The algorithm was employed to classify slides into two categories: benign (no further review needed) or suspicious (additional histologic and/or immunohistochemical analysis required). We assessed the sensitivity, specificity, positive predictive values (PPVs), and negative predictive values (NPVs) of a local pathologist, two central pathologists, and Paige Prostate in the diagnosis of 600 transrectal ultrasound‐guided prostate needle core biopsy regions (‘part‐specimens’) from 100 consecutive patients, and to ascertain the impact of Paige Prostate on diagnostic accuracy and efficiency. Paige Prostate displayed high sensitivity (0.99; CI 0.96–1.0), NPV (1.0; CI 0.98–1.0), and specificity (0.93; CI 0.90–0.96) at the part‐specimen level. At the patient level, Paige Prostate displayed optimal sensitivity (1.0; CI 0.93–1.0) and NPV (1.0; CI 0.91–1.0) at a specificity of 0.78 (CI 0.64–0.89). The 27 part‐specimens considered by Paige Prostate as suspicious, whose final diagnosis was benign, were found to comprise atrophy (n = 14), atrophy and apical prostate tissue (n = 1), apical/benign prostate tissue (n = 9), adenosis (n = 2), and post‐atrophic hyperplasia (n = 1). Paige Prostate resulted in the identification of four additional patients whose diagnoses were upgraded from benign/suspicious to malignant. Additionally, this AI‐based test provided an estimated 65.5% reduction of the diagnostic time for the material analyzed. Given its optimal sensitivity and NPV, Paige Prostate has the potential to be employed for the automated identification of patients whose histologic slides could forgo full histopathologic review. In addition to providing incremental improvements in diagnostic accuracy and efficiency, this AI‐based system identified patients whose prostate cancers were not initially diagnosed by three experienced histopathologists. © 2021 The Authors. The Journal of Pathology published by John Wiley & Sons, Ltd. on behalf of The Pathological Society of Great Britain and Ireland.

Abstract PD6-03: Clinical-grade detection of breast cancer in biopsies and excisions using machine learning

Background: Pathologists reviewing breast tissue slides must identify the presence of many salient features within each slide, including invasive and in situ breast cancer as well as various forms of atypia. In breast pathology particularly, the large volume of slides poses significant challenges for workload management and pathologist productivity (Johnson et al. 2019). The shift to a digital workflow in pathology, augmented by machine learning algorithms, has the potential to increase the efficiency, and productivity of pathologists by identifying cancer and pre-cancerous lesions in digitized slides. While there has been extensive work using machine learning algorithms to detect breast cancer metastasis in lymph nodes (Liu et al. 2018; Steiner et al. 2018), almost no research has been done on using such systems to detect breast cancer in biopsies and excisions.

Methods: We created and assessed Paige Breast Alpha, a machine learning system for the detection of breast cancer in hematoxylin and eosin (H&E) stained whole slide images (WSIs) of glass slides. The system is a binary classifier, intended to draw a pathologist’s attention to concerning features. Concerning features (the positive category) consisted of invasive breast cancer, in situ breast cancer, and various forms of atypia (atypical ductal hyperplasia, atypical lobular hyperplasia, etc. ). The deep learning system is based on the method proposed in Campanella et al. (2019). It learns directly from diagnosis using multiple instance learning, without the need for pixel-wise annotations. Paige Breast Alpha was trained on 17354 images from 3378 patients, and was assessed on 7921 images from 2443 patients. All slides were scanned on a Leica Aperio AT2.

Results: For detecting invasive or in situ cancer at the part level, the system achieved an overall sensitivity of 97.3% and a specificity of 98.0% in biopsies and 96.1% sensitivity and 91.5% specificity in excisions. Each part had between 1—10 slides.

Conclusions: We hypothesized that a machine learning system trained to detect predefined types of breast cancer and pre-cancerous lesions could be applied to a range of breast biopsy and resection WSIs to detect for the presence of these lesions. Herein we showed that the presence of these predefined features can be detected with high accuracy. Future studies are being initiated to assess the potential benefits of such a system when used by a pathologist. Additional systems are under development that would have the capability of subtyping the lesion present, in addition to acting as an overall binary classifier.

Citation Format: Matthew G Hanna, Patricia Raciti, Ran Godrich, Adam Casson, Julian Viret, Donghun Lee, Matthew Lee, Alican Bozkurt, Jillian Sue, Belma Dogdas, Brandon Rothrock, Leo Grady, Christopher Kanan, Thomas Fuchs. Clinical-grade detection of breast cancer in biopsies and excisions using machine learning [abstract]. In: Proceedings of the 2020 San Antonio Breast Cancer Virtual Symposium; 2020 Dec 8-11; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2021;81(4 Suppl):Abstract nr PD6-03.

Independent validation of paige prostate: Assessing clinical benefit of an artificial intelligence tool within a digital diagnostic pathology laboratory workflow

e14076

Background: The most common approach to diagnose prostate cancer is the “whole gland biopsy procedure,” in which numerous cores (≥12) are taken from different regions of the gland to maximize the chances of detecting small cancers; the presence of cancer in any of these cores is significant to the patient. If concerning features that are not fully diagnostic of cancer are identified, the pathologist may defer the final diagnosis until additional studies (e.g. immunohistochemistry) have been performed. We recently developed an artificial intelligence (AI)-based system for the assessment of cancer in prostate biopsies. Here, we investigated the performance of this test in an independent dataset of prostate cancers consecutively accrued. Methods: Two board-certified pathologists retrospectively reviewed 600 digitized hematoxylin-and-eosin (H&E) stained diagnostic prostate core needle biopsy slides from 100 consecutive patients, originally diagnosed at an independent hospital. Pathologists’ assessments were based on the H&E image alone; if further testing would be preferred, it was noted in the review notes. All images were assessed by Paige Prostate 1.0, an AI-based diagnostic tool; based on its outputs (either suspicious for cancer or not), the discordant images were re-reviewed by the pathologists and, in parallel, adjudicated with additional testing (e.g. ancillary immunohistochemical markers). Results: Paige Prostate's slide-level sensitivity was 98.9% and its specificity was 93.3% (100% and 78.0%, respectively, at the subject-level). The pathologists' average slide-level sensitivity and specificity without Paige Prostate was 90.9% and 98.6%, respectively. The sensitivity with their consensus read and Paige Prostate increased by 5.7% to 96.6% with only 0.8% decrease in specificity. In addition to new slide-level findings, benefits were also observed at the subject-level; with Paige, three new prostate cancer cases were discovered that were initially missed. Conclusions: The study reflects the potential benefits of the Paige Prostate system in the hands of experienced pathologists and validates the algorithm in a completely independent dataset. Paige Prostate can improve pathologists' sensitivity when reviewing digitized H&E prostate needle biopsy images with a minor impact on specificity.

Computational pathological identification of prostate cancer following neoadjuvant treatment

e14052

Background: The need for accurate pathological identification and quantitation of prostate cancer (PC) following neoadjuvant treatment with androgen deprivation therapy (ADT) and androgen receptor antagonists is increasing as PC treatment continues to evolve. In clinical practice, pathological assessment of residual tumor is a tedious and time-consuming process due to the volume of tissue from radical prostatectomy (RP). In addition, neoadjuvant treatments can greatly alter both benign and neoplastic prostate tissue morphology making the pathology assessment difficult for even specialized pathologists. Paige Prostate 1.0 is a clinical-grade artificial intelligence (AI) system for PC detection. It was trained and evaluated in over 50,000 prostate biopsy slides with validation across more than 800 institutions worldwide using multiple slide scanners. Methods: We evaluated the performance of Paige Prostate 1.0 at identifying prostatic tumor on 64 hematoxylin and eosin stained slides exhibiting neoadjuvant treatment effect from apalutamide, enzalutamide, and/or ADT. Results: Analysis of the receiver operating characteristic curve demonstrated an area under the curve of 0.96. Using the Paige Prostate 1.0 operating point, it achieved a sensitivity of 91% and a specificity of 94%, corresponding to the correct identification of challenging treated morphology in 59/64 slides using expert pathologists as the reference. False negative cases were typically represented by atypical small acinar proliferation that required expert pathological consensus confirmation. Conclusions: To our knowledge, this is the first AI based evaluation of residual disease in PC with hormone neoadjuvant therapy. Paige Prostate 1.0 effectively identified tumor despite treatment effects. Future work will include optimization of Paige Prostate 1.0 by training with RP specimens from a larger cohort of appropriate samples, as well as precise measurement of residual tumor burden to further improve its accuracy and reproducibility. Paige prostate residual disease detection 1.0 has the potential to impact emerging clinical practice at the patient level and to complement the pathological assessment of RPs in global phase 3 clinical trials, such as PROTEUS, in a standardized, reproducible, and robust way.

Precise measurement of coronary stenosis diameter with CCTA using CT number calibration

PURPOSE

Coronary x-ray computed tomography angiography (CCTA) continues to develop as a noninvasive method for the assessment of coronary vessel geometry and the identification of physiologically significant lesions. The uncertainty of quantitative lesion diameter measurement due to limited spatial resolution and vessel motion reduces the accuracy of CCTA diagnoses. In this paper, we introduce a new technique called computed tomography (CT)-number-Calibrated Diameter to improve the accuracy of the vessel and stenosis diameter measurements with CCTA.

METHODS

A calibration phantom containing cylindrical holes (diameters spanning from 0.8 mm through 4.0 mm) capturing the range of diameters found in human coronary vessels was three-dimensional printed. We also printed a human stenosis phantom with 17 tubular channels having the geometry of lesions derived from patient data. We acquired CT scans of the two phantoms with seven different imaging protocols. Calibration curves relating vessel intraluminal maximum voxel value (maximum CT number of a voxel, described in Hounsfield Units, HU) to true diameter, and full-width-at-half maximum (FWHM) to true diameter were constructed for each CCTA protocol. In addition, we acquired scans with a small constant motion (15 mm/s) and used a motion correction reconstruction (Snapshot Freeze) algorithm to correct motion artifacts. We applied our technique to measure the lesion diameter in the 17 lesions in the stenosis phantom and compared the performance of CT-number-Calibrated Diameter to the ground truth diameter and a FWHM estimate.

RESULTS

In all cases, vessel intraluminal maximum voxel value vs diameter was found to have a simple functional form based on the two-dimensional point spread function yielding a constant maximum voxel value region above a cutoff diameter, and a decreasing maximum voxel value vs decreasing diameter below a cutoff diameter. After normalization, focal spot size and reconstruction kernel were the principal determinants of cutoff diameter and the rate of maximum voxel value reduction vs decreasing diameter. The small constant motion had a significant effect on the CT number calibration; however, the motion-correction algorithm returned the maximum voxel value vs diameter curve to that of stationary vessels. The CT number Calibration technique showed better performance than FWHM estimation of diameter, yielding a high accuracy in the tested range (0.8 mm through 2.5 mm). We found a strong linear correlation between the smallest diameter in each of 17 lesions measured by CT-number-Calibrated Diameter (D_C ) and ground truth diameter (D_gt ), (D_C  = 0.951 × D_gt  + 0.023 mm, r = 0.998 with a slope very close to 1.0 and intercept very close to 0 mm.

CONCLUSIONS

Computed tomography-number-Calibrated Diameter is an effective method to enhance the accuracy of the estimate of small vessel diameters and degree of coronary stenosis in CCTA.

Automated estimation of image quality for coronary computed tomographic angiography using machine learning

OBJECTIVES

Our goal was to evaluate the efficacy of a fully automated method for assessing the image quality (IQ) of coronary computed tomography angiography (CCTA).

METHODS

The machine learning method was trained using 75 CCTA studies by mapping features (noise, contrast, misregistration scores, and un-interpretability index) to an IQ score based on manual ground truth data. The automated method was validated on a set of 50 CCTA studies and subsequently tested on a new set of 172 CCTA studies against visual IQ scores on a 5-point Likert scale.

RESULTS

The area under the curve in the validation set was 0.96. In the 172 CCTA studies, our method yielded a Cohen's kappa statistic for the agreement between automated and visual IQ assessment of 0.67 (p < 0.01). In the group where good to excellent (n = 163), fair (n = 6), and poor visual IQ scores (n = 3) were graded, 155, 5, and 2 of the patients received an automated IQ score > 50 %, respectively.

CONCLUSION

Fully automated assessment of the IQ of CCTA data sets by machine learning was reproducible and provided similar results compared with visual analysis within the limits of inter-operator variability.

KEY POINTS

• The proposed method enables automated and reproducible image quality assessment. • Machine learning and visual assessments yielded comparable estimates of image quality. • Automated assessment potentially allows for more standardised image quality. • Image quality assessment enables standardization of clinical trial results across different datasets.

Contrast Driven Elastica for Image Segmentation

Minimization of boundary curvature is a classic regularization technique for image segmentation in the presence of noisy image data. Techniques for minimizing curvature have historically been derived from gradient descent methods which could be trapped by a local minimum and, therefore, required a good initialization. Recently, combinatorial optimization techniques have overcome this barrier by providing solutions that can achieve a global optimum. However, curvature regularization methods can fail when the true object has high curvature. In these circumstances, existing methods depend on a data term to overcome the high curvature of the object. Unfortunately, the data term may be ambiguous in some images, which causes these methods also to fail. To overcome these problems, we propose a contrast driven elastica model (including curvature), which can accommodate high curvature objects and an ambiguous data model. We demonstrate that we can accurately segment extremely challenging synthetic and real images with ambiguous data discrimination, poor boundary contrast, and sharp corners. We provide a quantitative evaluation of our segmentation approach when applied to a standard image segmentation data set.

Fast Computation of Hemodynamic Sensitivity to Lumen Segmentation Uncertainty

Patient-specific blood flow modeling combining imaging data and computational fluid dynamics can aid in the assessment of coronary artery disease. Accurate coronary segmentation and realistic physiologic modeling of boundary conditions are important steps to ensure a high diagnostic performance. Segmentation of the coronary arteries can be constructed by a combination of automated algorithms with human review and editing. However, blood pressure and flow are not impacted equally by different local sections of the coronary artery tree. Focusing human review and editing towards regions that will most affect the subsequent simulations can significantly accelerate the review process. We define geometric sensitivity as the standard deviation in hemodynamics-derived metrics due to uncertainty in lumen segmentation. We develop a machine learning framework for estimating the geometric sensitivity in real time. Features used include geometric and clinical variables, and reduced-order models. We develop an anisotropic kernel regression method for assessment of lumen narrowing score, which is used as a feature in the machine learning algorithm. A multi-resolution sensitivity algorithm is introduced to hierarchically refine regions of high sensitivity so that we can quantify sensitivities to a desired spatial resolution. We show that the mean absolute error of the machine learning algorithm compared to 3D simulations is less than 0.01. We further demonstrate that sensitivity is not predicted simply by anatomic reduction but also encodes information about hemodynamics which in turn depends on downstream boundary conditions. This sensitivity approach can be extended to other systems such as cerebral flow, electro-mechanical simulations, etc.

FOCUSR: feature oriented correspondence using spectral regularization--a method for precise surface matching

Existing methods for surface matching are limited by the tradeoff between precision and computational efficiency. Here, we present an improved algorithm for dense vertex-to-vertex correspondence that uses direct matching of features defined on a surface and improves it by using spectral correspondence as a regularization. This algorithm has the speed of both feature matching and spectral matching while exhibiting greatly improved precision (distance errors of 1.4 percent). The method, FOCUSR, incorporates implicitly such additional features to calculate the correspondence and relies on the smoothness of the lowest-frequency harmonics of a graph Laplacian to spatially regularize the features. In its simplest form, FOCUSR is an improved spectral correspondence method that nonrigidly deforms spectral embeddings. We provide here a full realization of spectral correspondence where virtually any feature can be used as an additional information using weights on graph edges, but also on graph nodes and as extra embedded coordinates. As an example, the full power of FOCUSR is demonstrated in a real-case scenario with the challenging task of brain surface matching across several individuals. Our results show that combining features and regularizing them in a spectral embedding greatly improves the matching precision (to a submillimeter level) while performing at much greater speed than existing methods.

Sparsity-promoting calibration for GRAPPA accelerated parallel MRI reconstruction

The amount of calibration data needed to produce images of adequate quality can prevent auto-calibrating parallel imaging reconstruction methods like generalized autocalibrating partially parallel acquisitions (GRAPPA) from achieving a high total acceleration factor. To improve the quality of calibration when the number of auto-calibration signal (ACS) lines is restricted, we propose a sparsity-promoting regularized calibration method that finds a GRAPPA kernel consistent with the ACS fit equations that yields jointly sparse reconstructed coil channel images. Several experiments evaluate the performance of the proposed method relative to unregularized and existing regularized calibration methods for both low-quality and underdetermined fits from the ACS lines. These experiments demonstrate that the proposed method, like other regularization methods, is capable of mitigating noise amplification, and in addition, the proposed method is particularly effective at minimizing coherent aliasing artifacts caused by poor kernel calibration in real data. Using the proposed method, we can increase the total achievable acceleration while reducing degradation of the reconstructed image better than existing regularized calibration methods.

Network, anatomical, and non-imaging measures for the prediction of ADHD diagnosis in individual subjects

Brain imaging methods have long held promise as diagnostic aids for neuropsychiatric conditions with complex behavioral phenotypes such as Attention-Deficit/Hyperactivity Disorder. This promise has largely been unrealized, at least partly due to the heterogeneity of clinical populations and the small sample size of many studies. A large, multi-center dataset provided by the ADHD-200 Consortium affords new opportunities to test methods for individual diagnosis based on MRI-observable structural brain attributes and functional interactions observable from resting-state fMRI. In this study, we systematically calculated a large set of standard and new quantitative markers from individual subject datasets. These features (>12,000 per subject) consisted of local anatomical attributes such as cortical thickness and structure volumes, and both local and global resting-state network measures. Three methods were used to compute graphs representing interdependencies between activations in different brain areas, and a full set of network features was derived from each. Of these, features derived from the inverse of the time series covariance matrix, under an L1-norm regularization penalty, proved most powerful. Anatomical and network feature sets were used individually, and combined with non-imaging phenotypic features from each subject. Machine learning algorithms were used to rank attributes, and performance was assessed under cross-validation and on a separate test set of 168 subjects for a variety of feature set combinations. While non-imaging features gave highest performance in cross-validation, the addition of imaging features in sufficient numbers led to improved generalization to new data. Stratification by gender also proved to be a fruitful strategy to improve classifier performance. We describe the overall approach used, compare the predictive power of different classes of features, and describe the most impactful features in relation to the current literature.

Random walks based multi-image segmentation: Quasiconvexity results and GPU-based solutions

We recast the Cosegmentation problem using Random Walker (RW) segmentation as the core segmentation algorithm, rather than the traditional MRF approach adopted in the literature so far. Our formulation is similar to previous approaches in the sense that it also permits Cosegmentation constraints (which impose consistency between the extracted objects from ≥ 2 images) using a nonparametric model. However, several previous nonparametric cosegmentation methods have the serious limitation that they require adding one auxiliary node (or variable) for every pair of pixels that are similar (which effectively limits such methods to describing only those objects that have high entropy appearance models). In contrast, our proposed model completely eliminates this restrictive dependence -the resulting improvements are quite significant. Our model further allows an optimization scheme exploiting quasiconvexity for model-based segmentation with no dependence on the scale of the segmented foreground. Finally, we show that the optimization can be expressed in terms of linear algebra operations on sparse matrices which are easily mapped to GPU architecture. We provide a highly specialized CUDA library for Cosegmentation exploiting this special structure, and report experimental results showing these advantages.

Global error minimization in image mosaicing using graph connectivity and its applications in microscopy

Several applications such as multiprojector displays and microscopy require the mosaicing of images (tiles) acquired by a camera as it traverses an unknown trajectory in 3D space. A homography relates the image coordinates of a point in each tile to those of a reference tile provided the 3D scene is planar. Our approach in such applications is to first perform pairwise alignment of the tiles that have imaged common regions in order to recover a homography relating the tile pair. We then find the global set of homographies relating each individual tile to a reference tile such that the homographies relating all tile pairs are kept as consistent as possible. Using these global homographies, one can generate a mosaic of the entire scene. We derive a general analytical solution for the global homographies by representing the pair-wise homographies on a connectivity graph. Our solution can accommodate imprecise prior information regarding the global homographies whenever such information is available. We also derive equations for the special case of translation estimation of an X-Y microscopy stage used in histology imaging and present examples of stitched microscopy slices of specimens obtained after radical prostatectomy or prostate biopsy. In addition, we demonstrate the superiority of our approach over tree-structured approaches for global error minimization.

Denoising sparse images from GRAPPA using the nullspace method

To accelerate magnetic resonance imaging using uniformly undersampled (nonrandom) parallel imaging beyond what is achievable with generalized autocalibrating partially parallel acquisitions (GRAPPA) alone, the DEnoising of Sparse Images from GRAPPA using the Nullspace method is developed. The trade-off between denoising and smoothing the GRAPPA solution is studied for different levels of acceleration. Several brain images reconstructed from uniformly undersampled k-space data using DEnoising of Sparse Images from GRAPPA using the Nullspace method are compared against reconstructions using existing methods in terms of difference images (a qualitative measure), peak-signal-to-noise ratio, and noise amplification (g-factors) as measured using the pseudo-multiple replica method. Effects of smoothing, including contrast loss, are studied in synthetic phantom data. In the experiments presented, the contrast loss and spatial resolution are competitive with existing methods. Results for several brain images demonstrate significant improvements over GRAPPA at high acceleration factors in denoising performance with limited blurring or smoothing artifacts. In addition, the measured g-factors suggest that DEnoising of Sparse Images from GRAPPA using the Nullspace method mitigates noise amplification better than both GRAPPA and L1 iterative self-consistent parallel imaging reconstruction (the latter limited here by uniform undersampling).

Power Watershed: A Unifying Graph-Based Optimization Framework

In this work, we extend a common framework for graph-based image segmentation that includes the graph cuts, random walker, and shortest path optimization algorithms. Viewing an image as a weighted graph, these algorithms can be expressed by means of a common energy function with differing choices of a parameter q acting as an exponent on the differences between neighboring nodes. Introducing a new parameter p that fixes a power for the edge weights allows us to also include the optimal spanning forest algorithm for watershed in this same framework. We then propose a new family of segmentation algorithms that fixes p to produce an optimal spanning forest but varies the power q beyond the usual watershed algorithm, which we term the power watershed. In particular, when q=2, the power watershed leads to a multilabel, scale and contrast invariant, unique global optimum obtained in practice in quasi-linear time. Placing the watershed algorithm in this energy minimization framework also opens new possibilities for using unary terms in traditional watershed segmentation and using watershed to optimize more general models of use in applications beyond image segmentation.

Combining registration and minimum surfaces for the segmentation of the left ventricle in cardiac cine MR images

This paper describes a system to automatically segment the left ventricle in all slices and all phases of cardiac cine magnetic resonance datasets. After localizing the left ventricle blood pool using motion, thresholding and clustering, slices are segmented sequentially. For each slice, deformable registration is used to align all the phases, candidates contours are recovered in the average image using shortest paths, and a minimal surface is built to generate the final contours. The advantage of our method is that the resulting contours follow the edges in each phase and are consistent over time. We demonstrate using 19 patient examples that the results are very good. The RMS distance between ground truth and our segmentation is only 1.6 pixels (2.7 mm) and the Dice coefficient is 0.89.

Minimal surfaces extend shortest path segmentation methods to 3D

Shortest paths have been used to segment object boundaries with both continuous and discrete image models. Although these techniques are well defined in 2D, the character of the path as an object boundary is not preserved in 3D. An object boundary in three dimensions is a 2D surface. However, many different extensions of the shortest path techniques to 3D have been previously proposed in which the 3D object is segmented via a collection of shortest paths rather than a minimal surface, leading to a solution which bears an uncertain relationship to the true minimal surface. Specifically, there is no guarantee that a minimal path between points on two closed contours will lie on the minimal surface joining these contours. We observe that an elegant solution to the computation of a minimal surface on a cellular complex (e.g., a 3D lattice) was given by Sullivan [47]. Sullivan showed that the discrete minimal surface connecting one or more closed contours may be found efficiently by solving a Minimum-cost Circulation Network Flow (MCNF) problem. In this work, we detail why a minimal surface properly extends a shortest path (in the context of a boundary) to three dimensions, present Sullivan's solution to this minimal surface problem via an MCNF calculation, and demonstrate the use of these minimal surfaces on the segmentation of image data.

The piecewise smooth Mumford-Shah functional on an arbitrary graph

The Mumford-Shah functional has had a major impact on a variety of image analysis problems, including image segmentation and filtering, and, despite being introduced over two decades ago, it is still in widespread use. Present day optimization of the Mumford-Shah functional is predominated by active contour methods. Until recently, these formulations necessitated optimization of the contour by evolving via gradient descent, which is known for its overdependence on initialization and the tendency to produce undesirable local minima. In order to reduce these problems, we reformulate the corresponding Mumford-Shah functional on an arbitrary graph and apply the techniques of combinatorial optimization to produce a fast, low-energy solution. In contrast to traditional optimization methods, use of these combinatorial techniques necessitates consideration of the reconstructed image outside of its usual boundary, requiring additionally the inclusion of regularization for generating these values. The energy of the solution provided by this graph formulation is compared with the energy of the solution computed via traditional gradient descent-based narrow-band level set methods. This comparison demonstrates that our graph formulation and optimization produces lower energy solutions than the traditional gradient descent based contour evolution methods in significantly less time. Finally, we demonstrate the usefulness of the graph formulation to apply the Mumford-Shah functional to new applications such as point clustering and filtering of nonuniformly sampled images.

Weights and topology: a study of the effects of graph construction on 3D image segmentation

Graph-based algorithms have become increasingly popular for medical image segmentation. The fundamental process for each of these algorithms is to use the image content to generate a set of weights for the graph and then set conditions for an optimal partition of the graph with respect to these weights. To date, the heuristics used for generating the weighted graphs from image intensities have largely been ignored, while the primary focus of attention has been on the details of providing the partitioning conditions. In this paper we empirically study the effects of graph connectivity and weighting function on the quality of the segmentation results. To control for algorithm-specific effects, we employ both the Graph Cuts and Random Walker algorithms in our experiments.

An energy minimization approach to the data driven editing of presegmented images/volumes

Fully automatic, completely reliable segmentation in medical images is an unrealistic expectation with today's technology. However, many automatic segmentation algorithms may achieve a near-correct solution, incorrect only in a small region. For these situations, an interactive editing tool is required, ideally in 3D, that is usually left to a manual correction. We formulate the editing task as an energy minimization problem that may be solved with a modified version of either graph cuts or the random walker 3D segmentation algorithms. Both algorithms employ a seeded user interface, that may be used in this scenario for a user to seed erroneous voxels as belonging to the foreground or the background. In our formulation, it is unnecessary for the user to specify both foreground and background seeds.

Random walks for image segmentation

A novel method is proposed for performing multilabel, interactive image segmentation. Given a small number of pixels with user-defined (or predefined) labels, one can analytically and quickly determine the probability that a random walker starting at each unlabeled pixel will first reach one of the prelabeled pixels. By assigning each pixel to the label for which the greatest probability is calculated, a high-quality image segmentation may be obtained. Theoretical properties of this algorithm are developed along with the corresponding connections to discrete potential theory and electrical circuits. This algorithm is formulated in discrete space (i.e., on a graph) using combinatorial analogues of standard operators and principles from continuous potential theory, allowing it to be applied in arbitrary dimension on arbitrary graphs.

Isoperimetric graph partitioning for image segmentation

Spectral graph partitioning provides a powerful approach to image segmentation. We introduce an alternate idea that finds partitions with a small isoperimetric constant, requiring solution to a linear system rather than an eigenvector problem. This approach produces the high quality segmentations of spectral methods, but with improved speed and stability.

Accurate Banded Graph Cut Segmentation of Thin Structures Using Laplacian Pyramids

The Graph Cuts method of interactive segmentation has become very popular in recent years. This method performs at interactive speeds for smaller images/volumes, but an unacceptable amount of storage and computation time is required for the large images/volumes common in medical applications. The Banded Graph Cut (BGC) algorithm was proposed to drastically increase the computational speed of Graph Cuts, but is limited to the segmentation of large, roundish objects. In this paper, we propose a modification of BGC that uses the information from a Laplacian pyramid to include thin structures into the band. Therefore, we retain the computational efficiency of BGC while providing quality segmentations on thin structures. We make quantitative and qualitative comparisons with BGC on images containing thin objects. Additionally, we show that the new parameter introduced in our modification provides a smooth transition from BGC to traditional Graph Guts.

Random walks for interactive organ segmentation in two and three dimensions: implementation and validation

A new approach to interactive segmentation based on random walks was recently introduced that shows promise for allowing physicians more flexibility to segment arbitrary objects in an image. This report has two goals: To introduce a novel computational method for applying the random walker algorithm in 2D/3D using the Graphics Processing Unit (GPU) and to provide quantitative validation studies of this algorithm relative to different targets, imaging modalities and interaction strategies.

Multiple-Year Experience in the Diagnosis of Viral Central Nervous System Infections with a Panel of Polymerase Chain Reaction Assays for Detection of 11 Viruses

BACKGROUND

Polymerase chain reaction (PCR) is becoming more common in diagnostic laboratories. In some instances, its value has been established. In other cases, assays exist, but their beneficial use has not been determined. This article summarizes findings from 3485 patients who underwent testing over a 6-year period in our laboratory.

METHODS

A panel of PCR assays was used for the detection of a range of viruses associated with central nervous system (CNS) infections. PCR results were analyzed in conjunction with information about patient age and sex, the time between onset and specimen collection, and other variables. Medical chart review was conducted for 280 patients to gain diagnostic and epidemiologic insight with regard to cases of unresolved encephalitis.

RESULTS

A total of 498 PCR-positive samples (14.3%) were detected. Enteroviruses accounted for the largest number (360 [72.3%]) of positive PCR results, followed by herpes simplex virus (76 [15.3%]), varicella-zoster virus (29 [5.82%]), and West Nile virus (WNV) (18 [3.61%]). Of 360 patients who tested positive for enterovirus, only 46 met the Centers for Disease Control and Prevention's encephalitis definition. It resulted in the greatest decrease (87.2%) in positive PCR results. Overall, the PCR positivity rate for specimens collected within 5 days after illness onset was 17.2%, compared with 8.6% for specimens collected > or =6 days after onset.

CONCLUSIONS

The value of PCR in the diagnosis of viral infections has been established. PCR is of lower value in the detection of WNV in CNS, compared with serological testing, but is of greater value in the detection of other arboviruses, particularly viruses in the California serogroup. Medical chart reviews indicated that apparent CNS infection resolves in approximately 50% of cases.

Compliance with hand hygiene practice in pediatric intensive care

OBJECTIVE

To determine the frequency and predictors of compliance with hand hygiene (HH) practice in pediatric intensive care.

DESIGN

Observational, prospective cohort study performed from February to April 2000.

SETTING

Three intensive care units at a tertiary care children's hospital.

PARTICIPANTS

Nurses, physicians, respiratory therapists, and other healthcare workers.

METHODS

During 156 30-min daytime observation periods, an unidentified observer monitored 2811 opportunities for HH during patient care and recorded HH compliance.

MEASUREMENTS AND MAIN RESULTS

Average HH compliance was 34% (946/2811). It was higher (p < 0.001) among respiratory therapists (68%; 171/251) than physicians (37%; 157/426) or nurses (29%; 587/2031). Contact with body fluid secretions was associated with the highest compliance (77%; 46/60), and contact with wounds (71%; 10/14) or indwelling devices (66%; 110/167) were associated with somewhat lesser compliance. The following were important predictors of compliance (all p < 0.01): being a respiratory therapist (odds ratio [OR], 5.1); working in the neonatal intensive care unit (OR, 1.6); and contact with invasive devices (OR, 2.5), wounds (OR, 6.9), or body fluids (OR, 11.5). Compliance was lowest after interrupted patient-care activities (9%; OR, 0.15). Surprisingly, decreased patient-to-nurse ratio (mean, 1.3 +/- 0.3) or opportunities per hr of care (mean, 37 +/- 7) were not independent predictors of compliance.

CONCLUSIONS

Average HH compliance was low, but it increased during high-risk patient-care activities. Intensified efforts are necessary to increase caretakers' compliance and the awareness of the risk of bacterial contamination after interrupted patient-care activities.

Disruption of neurotransmission in Drosophila mushroom body blocks retrieval but not acquisition of memory

Surgical, pharmacological and genetic lesion studies have revealed distinct anatomical sites involved with different forms of learning. Studies of patients with localized brain damage and work in rodent model systems, for example, have shown that the hippocampal formation participates in acquisition of declarative tasks but is not the site of their long-term storage. Such lesions are usually irreversible, however, which has limited their use for dissecting the temporal processes of acquisition, storage and retrieval of memories. Studies in bees and flies have similarly revealed a distinct anatomical region of the insect brain, the mushroom body, that is involved specifically in olfactory associative learning. We have used a temperature-sensitive dynamin transgene, which disrupts synaptic transmission reversibly and on the time-scale of minutes, to investigate the temporal requirements for ongoing neural activity during memory formation. Here we show that synaptic transmission from mushroom body neurons is required during memory retrieval but not during acquisition or storage. We propose that the hebbian processes underlying olfactory associative learning reside in mushroom body dendrites or upstream of the mushroom body and that the resulting alterations in synaptic strength modulate mushroom body output during memory retrieval.

nalyot, a mutation of the Drosophila myb-related Adf1 transcription factor, disrupts synapse formation and olfactory memory

nalyot (nal) is a novel olfactory memory mutant of Drosophila, encoding Adf1, a myb-related transcription factor. Following extended training sessions, Adf1 mutants show normal early memory but defective longterm memory. Adf1 shows widespread spatiotemporal expression, yet mutant alleles reveal no discernible disruptions in gross morphology of the nervous system. Studies at the larval neuromuscular junction, however, reveal a role for Adf1 in the modulation of synaptic growth-in contrast to the role established for dCREB2 in the control of synaptic function (Davis et al., 1996). These findings suggest that Adf1 and dCREB2 regulate distinct transcriptional cascades involved in terminal stages of synapse maturation. More generally, Adf1 provides a novel link between molecular mechanisms of developmental and behavioral plasticity.

Diagnosis of Jamestown Canyon encephalitis by polymerase chain reaction

In recent years, polymerase chain reaction (PCR) has been under study as a potential technique to improve the accuracy of diagnosis of suspected central nervous system viral infections. We describe a case of severe encephalitis in a previously healthy 20-year-old woman from New York who presented with headache, fever, and photophobia. Her illness was characterized by progressive worsening of her neurological status, leading to confusion, delirium, and status epilepticus. The diagnosis of Jamestown Canyon encephalitis was established by positive reverse transcriptase (RT)-PCR and nucleic acid sequencing of the band from both cerebrospinal fluid and brain tissue. The nucleotide sequence and the deduced amino acid sequence of the Jamestown Canyon virus from this patient were very similar to Jamestown Canyon virus isolates from mosquito pools in New York. This report suggests that RT-PCR assays could be important tools in the diagnostic workup of cases of encephalitis.

Identification of a novel group of bacteria in sludge from a deteriorated biological phosphorus removal reactor

The microbial diversity of a deteriorated biological phosphorus removal reactor was investigated by methods not requiring direct cultivation. The reactor was fed with media containing acetate and high levels of phosphate (P/C weight ratio, 8:100) but failed to completely remove phosphate in the effluent and showed very limited biological phosphorus removal activity. Denaturing gradient gel electrophoresis (DGGE) of PCR-amplified 16S ribosomal DNA was used to investigate the bacterial diversity. Up to 11 DGGE bands representing at least 11 different sequence types were observed; DNA from the 6 most dominant of these bands was further isolated and sequenced. Comparative phylogenetic analysis of the partial 16S rRNA sequences suggested that one sequence type was affiliated with the alpha subclass of the Proteobacteria, one was associated with the Legionella group of the gamma subclass of the Proteobacteria, and the remaining four formed a novel group of the gamma subclass of the Proteobacteria with no close relationship to any previously described species. The novel group represented approximately 75% of the PCR-amplified DNA, based on the DGGE band intensities. Two oligonucleotide rRNA probes for this novel group were designed and used in a whole-cell hybridization analysis to investigate the abundance of this novel group in situ. The bacteria were coccoid and 3 to 4 microm in diameter and represented approximately 35% of the total population, suggesting a relatively close agreement with the results obtained by the PCR-based DGGE method. Further, based on electron microscopy and standard staining microscopic analysis, this novel group was able to accumulate granule inclusions, possibly consisting of polyhydroxyalkanoate, inside the cells.

Evidence that fatal human infections with La Crosse virus may be associated with a narrow range of genotypes

La Crosse (LAC) virus belongs to the California (CAL) serogroup of the genus Bunyavirus, family Bunyaviridae. It is considered one of the most important mosquito-borne pathogens in North America, especially in the upper Mid-West, where it is associated with encephalitis during the time of year when mosquitoes are active. Infections occur most frequently in children and young adults and, while most cases are resolved after a period of intense illness, a small fraction (< 1%) are fatal. At present there have only been three isolates of LAC virus from humans all made from brain tissue postmortem. The cases yielding viruses are separated chronologically by 33 years and geographically from Minnesota/Wisconsin (1960, 1978) to Missouri (1993). The M RNA sequence of the first two isolates was previously reported. The present study extends the observations to the isolate from the 1993 case and includes several mosquito isolates as well. A comparison of the M RNAs of these viruses shows that for the human isolates both nucleotide sequence and the deduced amino-acid sequence of the encoded proteins are highly conserved, showing a maximum variation of only 0.91% and 0.69%, respectively. This high degree of conservation over time and space leads to the hypothesis that human infections with this particular genotype of LAC virus are those most likely to have a fatal outcome. It is also shown that a virus with this genotype could be found circulating in mosquitoes in an area more or less intermediate between the locations of the first and second fatal cases.

An innovative approach to respiratory management for long-term care patients

Chronic obstructive pulmonary disease (COPD) is a condition that incapacitates a person physically, psychologically, and socially. Most of the supportive literature of respiratory care programs addresses noninstitutionalized patients. Many long-term care patients have relinquished much of their responsibility for self-care and health promotion. Patients often lack knowledge of the meaning of their disease, the role of exercise, nutrition, smoking, and medication. These factors, coupled with long-term institutionalization, contribute to progressively increasing disability. Our COPD program builds on the concepts of enablement and excess disability. A multidisciplinary team provides patients with the "tools" that allow them to take active responsibility for their health.

The SV40 pseudovirus: its potential for general transduction in animal cells

The rates of reassociation of DNA from a monkey cell line (Vero), from SV40 pseudovirus (consisting of Vero DNA and virus protein coat), and from mature SV40 were measured. The results show that the host DNA in the pseudovirus contains repeated and unique sequences in the same proportions as normal host DNA; hence, no one portion of the host genome is preferentially incorporated in the pseudovirus. It was also shown that the Vero DNA in the pseudoviruses enters the nuclei of mouse embryo cells without loss of its physical integrity.