Exploiting Rules to Enhance Machine Learning in Extracting Information From Multi-Institutional Prostate Pathology Reports.

PURPOSE: Literature on clinical note mining has highlighted the superiority of machine learning (ML) over hand-crafted rules. Nevertheless, most studies assume the availability of large training sets, which is rarely the case. For this reason, in the clinical setting, rules are still common. We suggest 2 methods to leverage the knowledge encoded in pre-existing rules to inform ML decisions and obtain high performance, even with scarce annotations. METHODS: We collected 501 prostate pathology reports from 6 American hospitals. Reports were split into 2,711 core segments, annotated with 20 attributes describing the histology, grade, extension, and location of tumors. The data set was split by institutions to generate a cross-institutional evaluation setting. We assessed 4 systems, namely a rule-based approach, an ML model, and 2 hybrid systems integrating the previous methods: a Rule as Feature model and a Classifier Confidence model. Several ML algorithms were tested, including logistic regression (LR), support vector machine (SVM), and eXtreme gradient boosting (XGB). RESULTS: When training on data from a single institution, LR lags behind the rules by 3.5% (F1 score: 92.2% v 95.7%). Hybrid models, instead, obtain competitive results, with Classifier Confidence outperforming the rules by +0.5% (96.2%). When a larger amount of data from multiple institutions is used, LR improves by +1.5% over the rules (97.2%), whereas hybrid systems obtain +2.2% for Rule as Feature (97.7%) and +2.6% for Classifier Confidence (98.3%). Replacing LR with SVM or XGB yielded similar performance gains. CONCLUSION: We developed methods to use pre-existing handcrafted rules to inform ML algorithms. These hybrid systems obtain better performance than either rules or ML models alone, even when training data are limited.

Automatic Normalization of Anatomical Phrases in Radiology Reports Using Unsupervised Learning.

In today's radiology workflow, free-text reporting is established as the most common medium to capture, store, and communicate clinical information. Radiologists routinely refer to prior radiology reports of a patient to recall critical information for new diagnosis, which is quite tedious, time consuming, and prone to human error. Automatic structuring of report content is desired to facilitate such inquiry of information. In this work, we propose an unsupervised machine learning approach to automatically structure radiology reports by detecting and normalizing anatomical phrases based on the Systematized Nomenclature of Medicine-Clinical Terms (SNOMED CT) ontology. The proposed approach combines word embedding-based semantic learning with ontology-based concept mapping to derive the desired concept normalization. The word embedding model was trained using a large corpus of unlabeled radiology reports. Fifty-six anatomical labels were extracted from SNOMED CT as class labels of the whole human anatomy. The proposed framework was compared against a number of state-of-the-art supervised and unsupervised approaches. Radiology reports from three different clinical sites were manually labeled for testing. The proposed approach outperformed other techniques yielding an average precision of 82.6%. The proposed framework boosts the coverage and performance of conventional approaches for concept normalization, by applying word embedding techniques in semantic learning, while avoiding the challenge of having access to a large amount of annotated data, which is typically required for training classifiers.

Is the link between anatomical structure and function equally strong at all cognitive levels of processing?

Whereas low-level sensory processes can be linked to macroanatomy with great confidence, the degree to which high-level cognitive processes map onto anatomy is less clear. If function respects anatomy, more accurate intersubject anatomical registration should result in better functional alignment. Here, we use auditory functional magnetic resonance imaging and compare the effectiveness of affine and nonlinear registration methods for aligning anatomy and functional activation across subjects. Anatomical alignment was measured using normalized cross-correlation within functionally defined regions of interest. Functional overlap was assessed using t-statistics from the group analyses and the degree to which group statistics predict high and consistent signal change in individual data sets. In regions related to early stages of auditory processing, nonlinear registration resulted in more accurate anatomical registration and stronger functional overlap among subjects compared with affine. In frontal and temporal areas reflecting high-level processing of linguistic meaning, nonlinear registration also improved the accuracy of anatomical registration. However, functional overlap across subjects was not enhanced in these regions. Therefore, functional organization, relative to anatomy, is more variable in the frontal and temporal areas supporting meaning-based processes than in areas devoted to sensory/perceptual auditory processing. This demonstrates for the first time that functional variability increases systematically between regions supporting lower and higher cognitive processes.

Creating probabilistic maps of the face network in the adolescent brain: a multicentre functional MRI study.

Large-scale magnetic resonance (MR) studies of the human brain offer unique opportunities for identifying genetic and environmental factors shaping the human brain. Here, we describe a dataset collected in the context of a multi-centre study of the adolescent brain, namely the IMAGEN Study. We focus on one of the functional paradigms included in the project to probe the brain network underlying processing of ambiguous and angry faces. Using functional MR (fMRI) data collected in 1,110 adolescents, we constructed probabilistic maps of the neural network engaged consistently while viewing the ambiguous or angry faces; 21 brain regions responding to faces with high probability were identified. We were also able to address several methodological issues, including the minimal sample size yielding a stable location of a test region, namely the fusiform face area (FFA), as well as the effect of acquisition site (eight sites) and scanner (four manufacturers) on the location and magnitude of the fMRI response to faces in the FFA. Finally, we provided a comparison between male and female adolescents in terms of the effect sizes of sex differences in brain response to the ambiguous and angry faces in the 21 regions of interest. Overall, we found a stronger neural response to the ambiguous faces in several cortical regions, including the fusiform face area, in female (vs. male) adolescents, and a slightly stronger response to the angry faces in the amygdala of male (vs. female) adolescents.

A statistical model-based technique for accounting for prostate gland deformation in endorectal coil-based MR imaging.

In prostate brachytherapy procedures, combining high-resolution endorectal coil (ERC)-MRI with Computed Tomography (CT) images has shown to improve the diagnostic specificity for malignant tumors. Despite such advantage, there exists a major complication in fusion of the two imaging modalities due to the deformation of the prostate shape in ERC-MRI. Conventionally, nonlinear deformable registration techniques have been utilized to account for such deformation. In this work, we present a model-based technique for accounting for the deformation of the prostate gland in ERC-MR imaging, in which a unique deformation vector is estimated for every point within the prostate gland. Modes of deformation for every point in the prostate are statistically identified using a set of MR-based training set (with and without ERC-MRI). Deformation of the prostate from a deformed (ERC-MRI) to a non-deformed state in a different modality (CT) is then realized by first calculating partial deformation information for a limited number of points (such as surface points or anatomical landmarks) and then utilizing the calculated deformation from a subset of the points to determine the coefficient values for the modes of deformations provided by the statistical deformation model. Using a leave-one-out cross-validation, our results demonstrated a mean estimation error of 1mm for a MR-to-MR registration.

A validation framework for probabilistic maps using Heschl's gyrus as a model.

Probabilistic maps are useful in functional neuroimaging research for anatomical labeling and for data analysis. The degree to which a probability map can accurately estimate the location of a structure of interest in a new individual depends on many factors, including variability in the morphology of the structure of interest over subjects, the registration (normalization procedure and template) applied to align the brains among individuals for constructing a probability map, and the registration used to map a new subject's data set to the frame of the probabilistic map. Here, we take Heschl's gyrus (HG) as our structure of interest, and explore the impact of different registration methods on the accuracy with which a probabilistic map of HG can approximate HG in a new individual. We assess and compare the goodness of fit of probability maps generated using five different registration techniques, as well as evaluating the goodness of fit of a previously published probabilistic map of HG generated using affine registration (Penhune et al., 1996). The five registration techniques are: three groupwise registration techniques (implicit reference-based or IRG, DARTEL, and BSpline-based); a high-dimensional pairwise registration (HAMMER) as well as a segmentation-based registration (unified segmentation of SPM5). The accuracy of the resulting maps in labeling HG was assessed using evidence-based diagnostic measures within a leave-one-out cross-validation framework. Our results demonstrated the out performance of IRG and DARTEL compared to other registration techniques in terms of sensitivity, specificity and positive predictive value (PPV). All the techniques displayed relatively low sensitivity rates, despite high PPV, indicating that the generated probability maps provide accurate but conservative estimates of the location and extent of HG in new individuals.

Reducing inter-subject anatomical variation: effect of normalization method on sensitivity of functional magnetic resonance imaging data analysis in auditory cortex and the superior temporal region.

Conventional group analysis of functional MRI (fMRI) data usually involves spatial alignment of anatomy across participants by registering every brain image to an anatomical reference image. Due to the high degree of inter-subject anatomical variability, a low-resolution average anatomical model is typically used as the target template, and/or smoothing kernels are applied to the fMRI data to increase the overlap among subjects' image data. However, such smoothing can make it difficult to resolve small regions such as subregions of auditory cortex when anatomical morphology varies among subjects. Here, we use data from an auditory fMRI study to show that using a high-dimensional registration technique (HAMMER) results in an enhanced functional signal-to-noise ratio (fSNR) for functional data analysis within auditory regions, with more localized activation patterns. The technique is validated against DARTEL, a high-dimensional diffeomorphic registration, as well as against commonly used low-dimensional normalization techniques such as the techniques provided with SPM2 (cosine basis functions) and SPM5 (unified segmentation) software packages. We also systematically examine how spatial resolution of the template image and spatial smoothing of the functional data affect the results. Only the high-dimensional technique (HAMMER) appears to be able to capitalize on the excellent anatomical resolution of a single-subject reference template, and, as expected, smoothing increased fSNR, but at the cost of spatial resolution. In general, results demonstrate significant improvement in fSNR using HAMMER compared to analysis after normalization using DARTEL, or conventional normalization such as cosine basis function and unified segmentation in SPM, with more precisely localized activation foci, at least for activation in the region of auditory cortex.

A new approach for creating customizable cytoarchitectonic probabilistic maps without a template.

We present a novel technique for creating template-free probabilistic maps of the cytoarchitectonic areas using a groupwise registration. We use the technique to transform 10 human post-mortem structural MR data sets, together with their corresponding cytoarchitectonic information, to a common space. We have targeted the cytoarchitectonically defined subregions of the primary auditory cortex. Thanks to the template-free groupwise registration, the created maps are not macroanatomically biased towards a specific geometry/topology. The advantage of the group-wise versus pairwise registration in avoiding such anatomical bias is better revealed in studies with small number of subjects and a high degree of variability among the individuals such as the post-mortem data. A leave-one-out cross-validation method was used to compare the sensitivity, specificity and positive predictive value of the proposed and published maps. We observe a significant improvement in localization of cytoarchitectonically defined subregions in primary auditory cortex using the proposed maps. The proposed maps can be tailored to any subject space by registering the subject image to the average of the groupwise-registered post-mortem images.

A framework for the design of a novel haptic-based medical training simulator.

This paper presents a framework for the design of a haptic-based medical ultrasound training simulator. The proposed simulator is composed of a PHANToM haptic device and a modular software package that allows for visual feedback and kinesthetic interactions between an operator and multimodality image databases. The system provides real-time ultrasound images in the same fashion as a typical ultrasound machine, enhanced with corresponding augmented computerized tomographic (CT) and/or MRI images. The proposed training system allows trainees to develop radiology techniques and knowledge of the patient's anatomy with minimum practice on live patients, or in places or at times when radiology devices or patients with rare cases may not be available. Low-level details of the software structure that can be migrated to other similar medical simulators are described. A preliminary human factors study, conducted on the prototype of the developed simulator, demonstrates the potential usage of the system for clinical training.

A statistical atlas-based technique for automatic segmentation of the first Heschl's gyrus in human auditory cortex from MR images.

We present an automatic method for the segmentation of the first transverse temporal gyrus of Heschl (HG), the morphological marker for primary auditory cortex in humans. The proposed technique utilizes a statistical anatomical atlas of the gyrus, generated from a set of training samples using principal component analysis. The training set consists of MRI data from 12 subjects with the corresponding Heschl's gyri manually labeled in each hemisphere (separate atlases were generated for each hemisphere). We used a leave-one-out approach to automatically segment Heschl's gyri in both hemispheres from the MR image data using generated atlases. We assessed the accuracy of this atlas-based technique by using it to segment the HG region from several test cases and finding the overlap between the segmented and labeled HG regions. Results demonstrated more than 75% and 83% accuracy in the extraction of the HG volumes in the left and right hemispheres, respectively. It is expected that the proposed tool can be adapted to extract other anatomical regions in the brain.