Search

Rationale and Objectives: Studies that evaluate the lung nodule detection performance of radiologists or computerized methods depend on an initial inventory of the nodules within the thoracic images (the "truth"). The purpose of this study was to analyze (1) variability in the "truth" defined by different combinations of experienced thoracic radiologists and (2) variability in the performance of other experienced thoracic radiologists based on these definitions of "truth" in the context of lung nodule detection in computed tomographic (CT) scans.Materials and Methods: Twenty-five thoracic CT scans were reviewed by four thoracic radiologists, who independently marked lesions they considered to be nodules >or=3 mm in maximum diameter. Panel "truth" sets of nodules were then derived from the nodules marked by different combinations of two and three of these four radiologists. The nodule detection performance of the other radiologists was evaluated based on these panel "truth" sets.Results: The number of "true" nodules in the different panel "truth" sets ranged from 15 to 89 (mean 49.8 +/- 25.6). The mean radiologist nodule detection sensitivities across radiologists and panel "truth" sets for different panel "truth" conditions ranged from 51.0 to 83.2%; mean false-positive rates ranged from 0.33 to 1.39 per case.Conclusions: Substantial variability exists across radiologists in the task of lung nodule identification in CT scans. The definition of "truth" on which lung nodule detection studies are based must be carefully considered, because even experienced thoracic radiologists may not perform well when measured against the "truth" established by other experienced thoracic radiologists. [ABSTRACT FROM AUTHOR]

Objectives: Precision medicine requires the measurement, quantification, and cataloging of medical characteristics to identify the most effective medical intervention. However, the amount of available data exceeds our current capacity to extract meaningful information. We examine the informatics needs to achieve precision medicine from the perspective of quantitative imaging and oncology. Methods: The National Cancer Institute (NCI) organized several workshops on the topic of medical imaging and precision medicine. The observations and recommendations are summarized herein. Results: Recommendations include: use of standards in data collection and clinical correlates to promote interoperability; data sharing and validation of imaging tools; clinician's feedback in all phases of research and development; use of open-source architecture to encourage reproducibility and reusability; use of challenges which simulate real-world situations to incentivize innovation; partnership with industry to facilitate commercialization; and education in academic communities regarding the challenges involved with translation of technology from the research domain to clinical utility and the benefits of doing so. Conclusions: This article provides a survey of the role and priorities for imaging informatics to help advance quantitative imaging in the era of precision medicine. While these recommendations were drawn from oncology, they are relevant and applicable to other clinical domains where imaging aids precision medicine., Competing Interests: Disclosure The authors report no conflicts of interest in this work., (Georg Thieme Verlag KG Stuttgart.)

Imaging biomarkers (IBs) are integral to the routine management of patients with cancer. IBs used daily in oncology include clinical TNM stage, objective response and left ventricular ejection fraction. Other CT, MRI, PET and ultrasonography biomarkers are used extensively in cancer research and drug development. New IBs need to be established either as useful tools for testing research hypotheses in clinical trials and research studies, or as clinical decision-making tools for use in healthcare, by crossing 'translational gaps' through validation and qualification. Important differences exist between IBs and biospecimen-derived biomarkers and, therefore, the development of IBs requires a tailored 'roadmap'. Recognizing this need, Cancer Research UK (CRUK) and the European Organisation for Research and Treatment of Cancer (EORTC) assembled experts to review, debate and summarize the challenges of IB validation and qualification. This consensus group has produced 14 key recommendations for accelerating the clinical translation of IBs, which highlight the role of parallel (rather than sequential) tracks of technical (assay) validation, biological/clinical validation and assessment of cost-effectiveness; the need for IB standardization and accreditation systems; the need to continually revisit IB precision; an alternative framework for biological/clinical validation of IBs; and the essential requirements for multicentre studies to qualify IBs for clinical use.

The purpose of this work is to describe the LUNGx Challenge for the computerized classification of lung nodules on diagnostic computed tomography (CT) scans as benign or malignant and report the performance of participants' computerized methods along with that of six radiologists who participated in an observer study performing the same Challenge task on the same dataset. The Challenge provided sets of calibration and testing scans, established a performance assessment process, and created an infrastructure for case dissemination and result submission. Ten groups applied their own methods to 73 lung nodules (37 benign and 36 malignant) that were selected to achieve approximate size matching between the two cohorts. Area under the receiver operating characteristic curve (AUC) values for these methods ranged from 0.50 to 0.68; only three methods performed statistically better than random guessing. The radiologists' AUC values ranged from 0.70 to 0.85; three radiologists performed statistically better than the best-performing computer method. The LUNGx Challenge compared the performance of computerized methods in the task of differentiating benign from malignant lung nodules on CT scans, placed in the context of the performance of radiologists on the same task. The continued public availability of the Challenge cases will provide a valuable resource for the medical imaging research community.

The purpose of this editorial is to provide a brief history of National Institutes of Health National Cancer Institute (NCI) workshops as related to quantitative imaging within the oncology setting. The editorial will then focus on the recently supported NCI initiatives, including the Quantitative Imaging Network (QIN) initiative and its organizational structure, including planned research goals and deliverables. The publications in this issue of Translational Oncology come from many of the current members of this QIN research network.

Rationale and Objectives: Computer-aided diagnostic (CAD) systems fundamentally require the opinions of expert human observers to establish "truth" for algorithm development, training, and testing. The integrity of this "truth," however, must be established before investigators commit to this "gold standard" as the basis for their research. The purpose of this study was to develop a quality assurance (QA) model as an integral component of the "truth" collection process concerning the location and spatial extent of lung nodules observed on computed tomography (CT) scans to be included in the Lung Image Database Consortium (LIDC) public database., Materials and Methods: One hundred CT scans were interpreted by four radiologists through a two-phase process. For the first of these reads (the "blinded read phase"), radiologists independently identified and annotated lesions, assigning each to one of three categories: "nodule >or=3 mm," "nodule <3 mm," or "non-nodule >or=3 mm." For the second read (the "unblinded read phase"), the same radiologists independently evaluated the same CT scans, but with all of the annotations from the previously performed blinded reads presented; each radiologist could add to, edit, or delete their own marks; change the lesion category of their own marks; or leave their marks unchanged. The post-unblinded read set of marks was grouped into discrete nodules and subjected to the QA process, which consisted of identification of potential errors introduced during the complete image annotation process and correction of those errors. Seven categories of potential error were defined; any nodule with a mark that satisfied the criterion for one of these categories was referred to the radiologist who assigned that mark for either correction or confirmation that the mark was intentional., Results: A total of 105 QA issues were identified across 45 (45.0%) of the 100 CT scans. Radiologist review resulted in modifications to 101 (96.2%) of these potential errors. Twenty-one lesions erroneously marked as lung nodules after the unblinded reads had this designation removed through the QA process., Conclusions: The establishment of "truth" must incorporate a QA process to guarantee the integrity of the datasets that will provide the basis for the development, training, and testing of CAD systems.

Rationale and Objectives: The Lung Image Database Consortium (LIDC) is developing a publicly available database of thoracic computed tomography (CT) scans as a medical imaging research resource to promote the development of computer-aided detection or characterization of pulmonary nodules. To obtain the best estimate of the location and spatial extent of lung nodules, expert thoracic radiologists reviewed and annotated each scan. Because a consensus panel approach was neither feasible nor desirable, a unique two-phase, multicenter data collection process was developed to allow multiple radiologists at different centers to asynchronously review and annotate each CT scan. This data collection process was also intended to capture the variability among readers., Materials and Methods: Four radiologists reviewed each scan using the following process. In the first or "blinded" phase, each radiologist reviewed the CT scan independently. In the second or "unblinded" review phase, results from all four blinded reviews were compiled and presented to each radiologist for a second review, allowing the radiologists to review their own annotations together with the annotations of the other radiologists. The results of each radiologist's unblinded review were compiled to form the final unblinded review. An XML-based message system was developed to communicate the results of each reading., Results: This two-phase data collection process was designed, tested, and implemented across the LIDC. More than 500 CT scans have been read and annotated using this method by four expert readers; these scans either are currently publicly available at http://ncia.nci.nih.gov or will be in the near future., Conclusions: A unique data collection process was developed, tested, and implemented that allowed multiple readers at distributed sites to asynchronously review CT scans multiple times. This process captured the opinions of each reader regarding the location and spatial extent of lung nodules.

Rationale and Objectives: The goal was to investigate the effects of choosing between different metrics in estimating the size of pulmonary nodules as a factor both of nodule characterization and of performance of computer aided detection systems, because the latter are always qualified with respect to a given size range of nodules., Materials and Methods: This study used 265 whole-lung CT scans documented by the Lung Image Database Consortium (LIDC) using their protocol for nodule evaluation. Each inspected lesion was reviewed independently by four experienced radiologists who provided boundary markings for nodules larger than 3 mm. Four size metrics, based on the boundary markings, were considered: a unidimensional and two bidimensional measures on a single image slice and a volumetric measurement based on all the image slices. The radiologist boundaries were processed and those with four markings were analyzed to characterize the interradiologist variation, while those with at least one marking were used to examine the difference between the metrics., Results: The processing of the annotations found 127 nodules marked by all of the four radiologists and an extended set of 518 nodules each having at least one observation with three-dimensional sizes ranging from 2.03 to 29.4 mm (average 7.05 mm, median 5.71 mm). A very high interobserver variation was observed for all these metrics: 95% of estimated standard deviations were in the following ranges for the three-dimensional, unidimensional, and two bidimensional size metrics, respectively (in mm): 0.49-1.25, 0.67-2.55, 0.78-2.11, and 0.96-2.69. Also, a very large difference among the metrics was observed: 0.95 probability-coverage region widths for the volume estimation conditional on unidimensional, and the two bidimensional size measurements of 10 mm were 7.32, 7.72, and 6.29 mm, respectively., Conclusions: The selection of data subsets for performance evaluation is highly impacted by the size metric choice. The LIDC plans to include a single size measure for each nodule in its database. This metric is not intended as a gold standard for nodule size; rather, it is intended to facilitate the selection of unique repeatable size limited nodule subsets.

Rationale and Objectives: The purpose of this study was to analyze the variability of experienced thoracic radiologists in the identification of lung nodules on computed tomography (CT) scans and thereby to investigate variability in the establishment of the "truth" against which nodule-based studies are measured., Materials and Methods: Thirty CT scans were reviewed twice by four thoracic radiologists through a two-phase image annotation process. During the initial "blinded read" phase, radiologists independently marked lesions they identified as "nodule >or=3 mm (diameter)," "nodule <3 mm," or "non-nodule >or=3 mm." During the subsequent "unblinded read" phase, the blinded read results of all four radiologists were revealed to each radiologist, who then independently reviewed their marks along with the anonymous marks of their colleagues; a radiologist's own marks then could be deleted, added, or left unchanged. This approach was developed to identify, as completely as possible, all nodules in a scan without requiring forced consensus., Results: After the initial blinded read phase, 71 lesions received "nodule >or=3 mm" marks from at least one radiologist; however, all four radiologists assigned such marks to only 24 (33.8%) of these lesions. After the unblinded reads, a total of 59 lesions were marked as "nodule >or=3 mm" by at least one radiologist. Twenty-seven (45.8%) of these lesions received such marks from all four radiologists, three (5.1%) were identified as such by three radiologists, 12 (20.3%) were identified by two radiologists, and 17 (28.8%) were identified by only a single radiologist., Conclusion: The two-phase image annotation process yields improved agreement among radiologists in the interpretation of nodules >or=3 mm. Nevertheless, substantial variability remains across radiologists in the task of lung nodule identification.

Rationale and Objectives: Integral to the mission of the National Institutes of Health-sponsored Lung Imaging Database Consortium is the accurate definition of the spatial location of pulmonary nodules. Because the majority of small lung nodules are not resected, a reference standard from histopathology is generally unavailable. Thus assessing the source of variability in defining the spatial location of lung nodules by expert radiologists using different software tools as an alternative form of truth is necessary., Materials and Methods: The relative differences in performance of six radiologists each applying three annotation methods to the task of defining the spatial extent of 23 different lung nodules were evaluated. The variability of radiologists' spatial definitions for a nodule was measured using both volumes and probability maps (p-map). Results were analyzed using a linear mixed-effects model that included nested random effects., Results: Across the combination of all nodules, volume and p-map model parameters were found to be significant at P < .05 for all methods, all radiologists, and all second-order interactions except one. The radiologist and methods variables accounted for 15% and 3.5% of the total p-map variance, respectively, and 40.4% and 31.1% of the total volume variance, respectively., Conclusion: Radiologists represent the major source of variance as compared with drawing tools independent of drawing metric used. Although the random noise component is larger for the p-map analysis than for volume estimation, the p-map analysis appears to have more power to detect differences in radiologist-method combinations. The standard deviation of the volume measurement task appears to be proportional to nodule volume.

To stimulate the advancement of computer-aided diagnostic (CAD) research for lung nodules in thoracic computed tomography (CT), the National Cancer Institute launched a cooperative effort known as the Lung Image Database Consortium (LIDC). The LIDC is composed of five academic institutions from across the United States that are working together to develop an image database that will serve as an international research resource for the development, training, and evaluation of CAD methods in the detection of lung nodules on CT scans. Prior to the collection of CT images and associated patient data, the LIDC has been engaged in a consensus process to identify, address, and resolve a host of challenging technical and clinical issues to provide a solid foundation for a scientifically robust database. These issues include the establishment of (a) a governing mission statement, (b) criteria to determine whether a CT scan is eligible for inclusion in the database, (c) an appropriate definition of the term qualifying nodule, (d) an appropriate definition of "truth" requirements, (e) a process model through which the database will be populated, and (f) a statistical framework to guide the application of assessment methods by users of the database. Through a consensus process in which careful planning and proper consideration of fundamental issues have been emphasized, the LIDC database is expected to provide a powerful resource for the medical imaging research community. This article is intended to share with the community the breadth and depth of these key issues., (Copyright RSNA, 2004)

A group of experts on very large databases, quantitative imaging, data format standards development, image management and communications, and related technologies for cancer imaging met at a recent workshop sponsored by the BIP and discussed the key issues confronting this field. The BIP received recommendations regarding steps that can be taken to advance the technology and take advantage of the opportunities to improve collaboration and utility in cancer imaging. There are tremendous opportunities to change radically the way we use image information. These opportunities are most obvious in clinical research, in which we seek to advance the dissemination and use of cancer image data in research and practice. Important opportunities and new modes of information use are provided by supporting the entire "information cycle" of creation, dissemination, and collaboration in addition to image organization, access, presentation, and preservation. To learn more, visit the BIP Web site at www3 .cancer.gov/bip/steer&#95;iasc.htm and join the image archive listserver at the NIH (ARCHIVE-COMM-L, available online at list.nihgov). These resources are open and available to all.

Preliminary clinical studies suggest that spiral computed tomography (CT) of the lungs can improve early detection of lung cancer in high-risk individuals. More clinical studies are needed, however, before public health recommendations can be proposed for population-based screening. Spiral CT generates large-volume data sets and thus poses problems in terms of implementation of efficient and cost-effective screening methods. Image processing algorithms such as computer assisted diagnostic (CAD) methods have the potential to assist in lesion (eg, nodule) detection on spiral CT studies. CAD methods may also be used to characterize nodules by either assessing the stability or change in size of lesions based on evaluation of serial CT studies, or quantitatively measuring the temporal parameters related to contrast dynamics when using contrast material-enhanced CT studies. CAD methods therefore have the potential to enhance the sensitivity and specificity of spiral CT lung screening studies. Lung cancer screening studies now under investigation create an opportunity to develop an image database that will allow comparison and optimization of CAD algorithms. This database could serve as an important national resource for the academic and industrial research community that is currently involved in the development of CAD methods. The National Cancer Institute request for applications (RFA) (CA-01-001) has already been announced (April 2000) to establish and support a consortium of academic centers to develop this database, the consortium to be referred to as the Lung Image Database Consortium (LIDC). This RFA is now closed. Five academic sites have been selected to be members of the LIDC, the first meeting of this consortium is planned for spring of 2001, and a public meeting is to be held in 2002. This report is abstracted from the previously published RFA to serve as an example of how an initiative is developed by the National Cancer Institute to support a research resource. For specific details of the RFA, please access the following Internet site: http://www. nci.nih.gov/bip/NCI-DIPinisumm.htm#a11.

We show that digitized mammograms can be considered as evolving from a simple process. A given image results from passing a random input field through a linear filtering operation, where the filter transfer function has a self-similar characteristic. By estimating the functional form of the filter and solving the corresponding filtering equation, the analysis shows that the input field gray value distribution and spectral content can be approximated with parametric methods. The work gives a simple explanation for the variegated image appearance and multimodal character of the gray value distribution common to mammograms. Using the image analysis as a guide, a simulated mammogram is generated that has many statistical characteristics of real mammograms. Additional benefits may follow from understanding the functional form of the filter in conjunction with the input field characteristics that include the approximate parametric description of mammograms, showing the distinction between homogeneously dense and nondense images, and the development of mass analysis methods.

Magnetic resonance images (MRIs) of the brain are segmented to measure the efficacy of treatment strategies for brain tumors. To date, no reproducible technique for measuring tumor size is available to the clinician, which hampers progress of the search for good treatment protocols. Many segmentation techniques have been proposed, but the representation (features) of the MRI data has received little attention. A genetic algorithm (GA) search was used to discover a feature set from multi-spectral MRI data. Segmentations were performed using the fuzzy c-means (FCM) clustering technique. Seventeen MRI data sets from five patients were evaluated. The GA feature set produces a more accurate segmentation. The GA fitness function that achieves the best results is the Wilks's lambda statistic when applied to FCM clusters. Compared to linear discriminant analysis, which requires class labels, the same or better accuracy is obtained by the features constructed from a GA search without class labels, allowing fully operator independent segmentation. The GA approach therefore provides a better starting point for the measurement of the response of a brain tumor to treatment.

Current MRI nonuniformity correction techniques are reviewed and investigated. Many approaches are used to remedy this artifact, but it is not clear which method is the most appropriate in a given situation, as the applications have been with different MRI coils and different clinical applications. In this work four widely used nonuniformity correction techniques are investigated in order to assess the effect on tumor response measurements (change in tumor volume over time): a phantom correction method, an image smoothing technique, homomorphic filtering, and surface fitting approach. Six brain tumor cases with baseline and follow-up MRIs after treatment with varying degrees of difficulty of segmentation were analyzed without and with each of the nonuniformity corrections. Different methods give significantly different correction images, indicating that rf nonuniformity correction is not yet well understood. No improvement in tumor segmentation or in tumor growth/shrinkage assessment was achieved using any of the evaluated corrections.

Rationale and Objectives: The authors evaluated an algorithm for the automatic segmentation of microcalcification clusters (MCCs) at digital mammography. Two- and four-channel wavelet transforms were evaluated to determine whether sensitivity in the detection of MCCs can be improved and if the selective reconstruction of the higher-order M2 subimages allows better preservation of the segmented MCCs, which is required for their classification., Materials and Methods: The hybrid method involved the use of a nonlinear filter for image noise suppression coupled with wavelet transforms for image decomposition and an adaptive method for selective subimage reconstruction as a basis for segmentation of MCCs. The two- and four-channel wavelet transforms were implemented with different filter bank structures (i.e., polyphase quadrature mirror filters [QMFs], tree structure, and lattice structure) to determine if their computational efficiency can be improved while retaining properties such as near-perfect reconstruction. The hybrid wavelet transforms were applied to a common image database of biopsy-proved MCCs (100 images, 105-micron resolution, 12 bits deep; 52 cases with at least one MCC of varying subtlety [46 malignant and six benign cases] and eight normal cases)., Results: The two- and four-channel wavelet transforms yielded sensitivities of 93% and 94% and false-positive (PP) detection rates of 1.58 and 1.35 MCCs per image, respectively. The lattice structure provided greater than fivefold improvement in computational speed compared to the polyphase QMF structure, particularly for the higher order of channels (M = 4)., Conclusion: The four-channel wavelet transform provided better sensitivity and FP detection rates and greater image detail preservation for the segmented MCCs.

Rationale and Objectives: The authors evaluated computer-assisted diagnostic (CAD) methods used to detect suspicious areas on lung radiographs., Materials and Methods: The authors designed a fragmentary window filtering (FWF) algorithm for detecting lung nodule patterns, which generally appear as circular areas of high opacity on the chest radiograph. The FWF algorithm helps differentiate circular patterns from overlapping radiographic background. A multiscale analysis was performed to locate multiscale nodules. Receiver operating characteristic analysis was performed by using a lung nodule that was extracted from a chest radiograph. The nodule underwent scalings and subsequent superimposition onto 140 normal regions of interest from six chest radiographs., Results: The FWF method was superior to the matched filtering method in the detection of suspicious areas., Conclusion: The proposed FWF-based method should provide improved detection of lung nodules on chest radiographs.

An automatic magnetic resonance imaging (MRI) multispectral segmentation method and a visual metric are compared for their effectiveness to measure tumor response to therapy. Automatic response measurements are important for multicenter clinical trials. A visual metric such as the product of the largest diameter and the largest perpendicular diameter of the tumor is a standard approach, and is currently used in the Radiation Treatment Oncology Group (RTOG) and the Eastern Cooperative Oncology Group (EGOG) clinical trials. In the standard approach, the tumor response is based on the percentage change in the visual metric and is categorized into cure, partial response, stable disease, or progression. Both visual and automatic methods are applied to six brain tumor cases (gliomas) of varying levels of segmentation difficulty. The analyzed data were serial multispectral MR images, collected using MR contrast enhancement. A fully automatic knowledge guided method (KG) was applied to the MRI multispectral data, while the visual metric was taken from the MRI films using the T1 gadolinium enhanced image, with repeat measurements done by two radiologists and two residents. Tumor measurements from both visual and automatic methods are compared to "ground truth," (GT) i.e., manually segmented tumor. The KG method was found to slightly overestimate tumor volume, but in a consistent manner, and the estimated tumor response compared very well to hand-drawn ground truth with a correlation coefficient of 0.96. In contrast, the visually estimated metric had a large variation between observers, particularly for difficult cases, where the tumor margins are not well delineated. The inter-observer variation for the measurement of the visual metric was only 16%, i.e., observers generally agreed on the lengths of the diameters. However, in 30% of the studied cases no consensus was found for the categorical tumor response measurement, indicating that the categories are very sensitive to variations in the diameter measurements. Moreover, the method failed to correctly identify the response in half of the cases. The data demonstrate that automatic 3D methods are clearly necessary for objective and clinically meaningful assessment of tumor volume in single or multicenter clinical trials.

Rationale and Objectives: The authors evaluated a modular computer-assisted diagnosis (CAD) method for mass detection that uses computation of features in three domains (gray level, morphology, and directional texture). Their objectives were to improve the sensitivity of detection and reduce the false-positive (FP) detection rate., Materials and Methods: The directional wavelet transform (DWT) method, which uses both multiorientation and multiresolution wavelet transforms to improve image preprocessing and segmentation of suspicious areas and to extract both morphologic and directional texture features, was evaluated with a previously reported image database containing 50 normal and 45 abnormal digitized screen-film mammograms. The mammograms contained all mass types and included 16 minimal cancers. This method was compared with the Markov random field (MRF) method to avoid issues related to case selection criteria. Free-response receiver operating characteristic curves were compared for both DWT and MRF methods., Results: For the DWT method, the sensitivity was 98% and the FP detection rate was 1.8 FP findings per image. For the MRF method, the sensitivity was 90% and the FP detection rate was 2.0 FP findings per image., Conclusion: The CAD method applied to the full mammographic image is automatic and independent of mass type. The segmentation of masses as performed with this method may potentially allow visual interpretation according to American College of Radiology criteria.

A multiresolution statistical method for identifying clinically normal tissue in digitized mammograms is used to construct an algorithm for separating normal regions from potentially abnormal regions; that is, small regions that may contain isolated calcifications. This is the initial phase of the development of a general method for the automatic recognition of normal mammograms. The first step is to decompose the image with a wavelet expansion that yields a sum of independent images, each containing different levels of image detail. When calcifications are present, there is strong empirical evidence that only some of the image components are necessary for the purpose of detecting a deviation from normal. The underlying statistic for each of the selected expansion components can be modeled with a simple parametric probability distribution function. This function serves as an instrument for the development of a statistical test that allows for the recognition of normal tissue regions. The distribution function depends on only one parameter, and this parameter itself has an underlying statistical distribution. The values of this parameter define a summary statistic that can be used to set detection error rates. Once the summary statistic is determined, spatial filters that are matched to resolution are applied independently to each selected expansion image. Regions of the image that correlate with the normal statistical model are discarded and regions in disagreement (suspicious areas) are flagged. These results are combined to produce a detection output image consisting only of suspicious areas. This type of detection output is amenable to further processing that may ultimately lead to a fully automated algorithm for the identification of normal mammograms.

This paper updates several recent surveys on the use of fuzzy models for segmentation and edge detection in medical image data. Our survey is divided into methods based on supervised and unsupervised learning (that is, on whether there are or are not labelled data available for supervising the computations), and is organized first and foremost by groups (that we know of!) that are active in this area. Our review is aimed more towards 'who is doing it' rather than 'how good it is'. This is partially dictated by the fact that direct comparisons of supervised and unsupervised methods is somewhat akin to comparing apples and oranges. There is a further subdivision into methods for two- and three-dimensional data and/or problems. We do not cover methods based on neural-like networks or fuzzy reasoning systems. These topics are covered in a recently published companion survey by keller et al.

The evaluation of algorithms' developed for computer assisted diagnosis in digital mammography requires image databases that allow relative comparisons and assessment of algorithms clinical value. A review of the literature indicates that there is no consensus on the guidelines of how databases should be established. Image selection is usually done based on subjective criteria or availability. The generation of common database(s) available to the research community makes relative evaluations of algorithms with similar properties easier. However, questions regarding the "right database size," the "right image resolution," and the "right contents" remain. In this paper, database issues are reviewed and discussed and possible remedies to the various problems are proposed.

The performance evaluation of a semi-supervised fuzzy c-means (SFCM) clustering method for monitoring brain tumor volume changes during the course of routine clinical radiation-therapeutic and chemo-therapeutic regimens is presented. The tumor volume determined using the SFCM method was compared with the volume estimates obtained using three other methods: (a) a k nearest neighbor (kNN) classifier, b) a grey level thresholding and seed growing (ISG-SG) method and c) a manual pixel labeling (GT) method for ground truth estimation. The SFCM and kNN methods are applied to the multispectral, contrast enhanced T1, proton density, and T2 weighted, magnetic resonance images (MRI) whereas the ISG-SG and GT methods are applied only to the contrast enhanced T1 weighted image. Estimations of tumor volume were made on eight patient cases with follow-up MRI scans performed over a 32 week interval during treatment. The tumor cases studied include one meningioma, two brain metastases and five gliomas. Comparisons with manually labeled ground truth estimations showed that there is a limited agreement between the segmentation methods for absolute tumor volume measurements when using images of patients after treatment. The average intraobserver reproducibility for the SFCM, kNN and ISG-SG methods was found to be 5.8%, 6.6% and 8.9%, respectively. The average of the interobserver reproducibility of these methods was found to be 5.5%, 6.5% and 11.4%, respectively. For the measurement of relative change of tumor volume as required for the response assessment, the multi-spectral methods kNN and SFCM are therefore preferred over the seedgrowing method.

The performance of a supervised k-nearest neighbor (kNN) classifier and a semisupervised fuzzy c-means (SFCM) clustering segmentation method are evaluated for reproducible measurement of the volumes of normal brain tissues and cerebrospinal fluid. The stability of the two segmentation methods is evaluated for (a) operator selection of training data, (b) reproducibility during repeat imaging sessions to determine any variations in the sensor performance over time, (c) variations in the measured volumes between different subjects, and (d) variability with different imaging parameters. The variations were found to be dependent on the type of measured tissue and the operator performing the segmentations. The variability during repeat imaging sessions for the SFCM method was < 3%. The absolute volumes of the brain matter and cerebrospinal fluid between subjects varied quite large, ranging from 9% to 13%. The intraobserver and interobserver reproducibility for SFCM were < 4% for the soft tissues and 6% for cerebrospinal fluid. The corresponding results for the kNN segmentation method were higher compared to the SFCM method.

A novel wavelet-based neural network (WNN) filter is proposed for image restoration as required for imaging of beta emitters by bremsstrahlung detection using a gamma camera. Quantitative imaging of beta emitters is important for the in vivo management of antibody therapy using either P-32 or Y-90. The theoretical basis for the general case for M-channel multiresolution wavelet decomposition of the nuclear image into different subimages is developed with the objective of isolating the signal from noise. A modified Hopfield neural network (NN) architecture is then used for multichannel image restoration using the dominant signal subimages. The NN model avoids the common inverse problem associated with other image restoration filters such as the Wiener filter. The relative performance of the WNN for image restoration, for M = 2 channel, is compared to a previously reported order statistic neural network hybrid (OSNNH) filter. Initially simulated degraded images of known structures with different noise levels are used. Quantitative metrics such as the normalized mean square error (NMSE) and signal-to-noise ratio (SNR) are used to compare filter performance. The WNN yields comparable results for image restoration with suggested slightly better performance for the images with higher noise levels as often encountered in bremsstrahlung detection. Attenuation measurements were performed using two radionuclides, 32P and 90Y as required for calibration of the gamma camera for quantitative measurements. Similar values for an effective attenuation coefficient were observed for the restored images using the OSNNH filters (32P: mu = 0.122 cm-1, 90Y: mu = 0.135 cm-1) and WNN (32P: mu = 0.122 cm-1, 90Y: mu = 0.135 cm-1) filters with slightly higher values obtained for the raw data (32P: mu = 0.142 cm-1, 90Y: mu = 0.142 cm-1) for a 3.5-cm source size. The WNN, however, was computationally more efficient by a factor of 4 to 6 compared to the OSNNH filter. The filter architecture, in turn, is also optimum for parallel processing or VLSI implementation as required for planar and particularly for SPECT mode of detection.

Rationale and Objectives: The acceptance of filmless digital mammography is currently limited by digitization and display drawbacks, as well as bias toward hard-copy interpretation. In the current study, we evaluated a wavelet-based image enhancement method for the filmless interpretation of breast calcifications., Methods: A set of 100 mammograms (58 with calcification clusters) was digitized at 105 microns and 4,096 gray levels per pixel and was processed with nonlinear filters and wavelets. Standard receiver operating characteristic analysis was performed by four radiologists, who independently read the films, the unprocessed digital images, and unprocessed and wavelet-enhanced digital images presented simultaneously., Results: Statistical differences were observed between screen/film and unprocessed digitized mammography displayed on monitors. Differences were not significant when wavelet enhancement was included in the monitor display. Interobserver variation in the digitized reading was greater than in film reading, but the wavelet enhancement reduced the difference., Conclusion: Wavelet-enhanced digital mammograms may assist radiologists in diagnosing calcifications directly from computer monitors and may compensate for current technologic limitations. A study with a larger data-base is needed before this method is accepted for clinical use.

A computationally efficient mixed feature based neural network (MFNN) is proposed for the detection of microcalcification clusters (MCCs) in digitized mammograms. The MFNN employs features computed in both the spatial and spectral domain and uses spectral entropy as a decision parameter. Backpropagation with Kalman filtering (KF) is employed to allow more efficient network training as required for evaluation of different features, input images, and related error analysis. A previously reported, wavelet-based image-enhancement method is also employed to enhance microcalcification clusters for improved detection. The relative performance of the MFNN for both the raw and enhanced images is evaluated using a common image database of 30 digitized mammograms, with 20 images containing 21 biopsy proven MCCs and ten normal cases. The computed sensitivity (true positive (TP) detection rate) was 90.1% with an average low false positive (FP) detection of 0.71 MCCs/image for the enhanced images using a modified k-fold validation error estimation technique. The corresponding computed sensitivity for the raw images was reduced to 81.4% and with 0.59 FP's MCCs/image. A relative comparison to an earlier neural network (NN) design, using only spatially related features, suggests the importance of the addition of spectral domain features when the raw image data is analyzed.

We examined unsupervised methods of segmentation of MR images of the brain for measuring tumor volume in response to treatment. Two clustering methods were used: fuzzy c-means and a nonfuzzy clustering algorithm. Results were compared with volume segmentations by two supervised methods, k-nearest neighbors and region growing, and all results were compared with manual labelings. Results of individual segmentations are presented as well as comparisons on the application of the different methods with 10 data sets of patients with brain tumors. Unsupervised segmentation is preferred for measuring tumor volumes in response to treatment, as it eliminates operator dependency and may be adequate for delineation of the target volume in radiation therapy. Some obstacles need to be overcome, in particular regarding the detection of anatomically relevant tissue classes. This study shows that these improvements are possible.

A novel multistage algorithm is proposed for the automatic segmentation of microcalcification clusters (MCCs) in digital mammography. First, a previously reported tree structured nonlinear filter is proposed for suppressing image noise, while preserving image details, to potentially reduce the false positive (FP) detection rate for MCCs. Second, a tree structured wavelet transform (TSWT) is applied to the images for microcalcification segmentation. The TSWT employs quadrature mirror filters as basic subunits for both multiresolution decomposition and reconstruction processes, where selective reconstruction of subimages is used to segment MCCs. Third, automatic linear scaling is then used to display the image of the segmented MCCs on a computer monitor for interpretation. The proposed algorithms were applied to an image database of 100 single view mammograms at a resolution of 105 microns and 12 bits deep (4096 gray levels). The database contained 50 cases of biopsy proven malignant MCCs, 8 benign cases, and 42 normal cases. The measured sensitivity (true positive detection rate) was 94% with a low FP detection rate of 1.6 MCCs/image. The image details of the segmented MCCs were reasonably well preserved, for microcalcification of less than 500 microns, with good delineation of the extent of the microcalcification clusters for each case based on visual criteria.

A technique is proposed for the detection of tumors in digital mammography. Detection is performed in two steps: segmentation and classification. In segmentation, regions of interest are first extracted from the images by adaptive thresholding. A further reliable segmentation is achieved by a modified Markov random field (MRF) model-based method. In classification, the MRF segmented regions are classified into suspicious and normal by a fuzzy binary decision tree based on a series of radiographic, density-related features. A set of normal (50) and abnormal (45) screen/film mammograms were tested. The latter contained 48 biopsy proven, malignant masses of various types and subtlety. The detection accuracy of the algorithm was evaluated by means of a free response receiver operating characteristic curve which shows the relationship between the detection of true positive masses and the number of false positive alarms per image. The results indicated that a 90% sensitivity can be achieved in the detection of different types of masses at the expense of two falsely detected signals per image. The algorithm was notably successful in the detection of minimal cancers manifested by masses

Two different multispectral pattern recognition methods are used to segment magnetic resonance images (MRI) of the brain for quantitative estimation of tumor volume and volume changes with therapy. A supervised k-nearest neighbor (kNN) rule and a semi-supervised fuzzy c-means (SFCM) method are used to segment MRI slice data. Tumor volumes as determined by the kNN and SFCM segmentation methods are compared with two reference methods, based on image grey scale, as a basis for an estimation of ground truth, namely: (a) a commonly used seed growing method that is applied to the contrast enhanced T1-weighted image, and (b) a manual segmentation method using a custom-designed graphical user interface applied to the same raw image (T1-weighted) dataset. Emphasis is placed on measurement of intra and inter observer reproducibility using the proposed methods. Intra- and interobserver variation for the kNN method was 9% and 5%, respectively. The results for the SFCM method was a little better at 6% and 4%, respectively. For the seed growing method, the intra-observer variation was 6% and the interobserver variation was 17%, significantly larger when compared with the multispectral methods. The absolute tumor volume determined by the multispectral segmentation methods was consistently smaller than that observed for the reference methods. The results of this study are found to be very patient case-dependent. The results for SFCM suggest that it should be useful for relative measurements of tumor volume during therapy, but further studies are required. This work demonstrates the need for minimally supervised or unsupervised methods for tumor volume measurements.

The application of a raw data-based, operator-independent MR segmentation technique to differentiate boundaries of tumor from edema or hemorrhage is demonstrated. A case of a glioblastoma multiforme with gross and histopathologic correlation is presented. The MR image data set was segmented into tissue classes based on three different MR weighted image parameters (T1-, proton density-, and T2-weighted) using unsupervised fuzzy c-means (FCM) clustering algorithm technique for pattern recognition. A radiological examination of the MR images and correlation with fuzzy clustering segmentations was performed. Results were confirmed by gross and histopathology which, to the best of our knowledge, reports the first application of this demanding approach. Based on the results of neuropathologic correlation, the application of FCM MR image segmentation to several MR images of a glioblastoma multiforme represents a viable technique for displaying diagnostically relevant tissue contrast information used in 3D volume reconstruction. With this technique, it is possible to generate segmentation images that display clinically important neuroanatomic and neuropathologic tissue contrast information from raw MR image data.

The current literature on MRI segmentation methods is reviewed. Particular emphasis is placed on the relative merits of single image versus multispectral segmentation, and supervised versus unsupervised segmentation methods. Image pre-processing and registration are discussed, as well as methods of validation. The application of MRI segmentation for tumor volume measurements during the course of therapy is presented here as an example, illustrating problems associated with inter- and intra-observer variations inherent to supervised methods.

Multiresolution methods are reported for feature extraction in breast cancer screening using digital mammography. The initial application is directed at the detection of microcalcification clusters (MCCs). Quadrature mirror filter (QMF) banks, using both two and three channel are proposed for the first time for both multiresolution decomposition and reconstruction. These filters are specifically tailored for automatic extraction of MCCs. The QMF multiresolution methods are compared to two channel tree structured wavelet transforms (TSWTs) methods previously reported. The QMF filters are preceded by an advanced tree structured nonlinear filter for noise suppression, prior to feature extraction, in order to minimize the false positive (FP) detection rate in digital mammography. The relative performance of these methods were evaluated using both simulated images and fifteen representative digitized mammograms containing biopsy proven microcalcification clusters. Similar high sensitivity (true positive (TP) detection rate (100%) and high specificity (0.6 average false positive (FP) MCC's/image) were observed, substantially better than more traditional approaches using single scale filters. The three channel QMF method, however, demonstrated better detail preservation of MCC's extracted compared to the two channel method. Detail preservation is important for the characterization of MCC's or individual microcalcifications in cancer screening.

As a first step in determining the efficacy of using computers to assist in diagnosis of medical images, an investigation has been conducted which utilizes the patterns, or textures, in the images. To be of value, any computer scheme must be able to recognize and differentiate the various patterns. An obvious example of this in mammography is the recognition of tumorous tissue and non-malignant abnormal tissue from normal parenchymal tissue. We have developed a pattern recognition technique which uses features derived from the fractal nature of the image. Further, we are able to develop mathematical models which can be used to differentiate and classify the many tissue types. Based on a limited number of cases of digitized mammograms, our computer algorithms have been able to distinguish tumorous from healthy tissue and to distinguish among various parenchymal tissue patterns. These preliminary results indicate that discrimination based on the fractal nature of images may well represent a viable approach to utilizing computers to assist in diagnosis.

A novel algorithm was developed for computer aided diagnosis of microcalcification clusters in digital mammography. The method includes: (a) tree-structured central weighted median filters with variable shape windowing to suppress image noise but preserve image details; (b) a quasi range dispersion edge detector to increase edge contrast and definition; and (c) tree-structured wavelets for calcification segmentation. The preliminary evaluation of the method on nine mammograms showed that 100% sensitivity can be achieved at the expense of four false positive clusters per image. Research is ongoing for further optimization of the algorithm to reduce the number of false alarms and establish its clinical value.

An initial evaluation of Haar wavelets is presented in this study for the compression of mammographic images. Fifteen mammograms with 105 microns/pixel resolution and varying dynamic range (10 and 12 bits per pixel) containing clustered microcalcifications were compressed with two different rates. The quality and content of the compressed reconstructed images was evaluated by an expert mammographer. The visualization of the cluster was on the average good but degraded with increasing compression because of the discontinuities introduced by these types of wavelets as the compression rate increases. However, the artifacts in the decoded images were seen as totally artificial and were not misinterpreted by the radiologist as calcifications. The classification of the parenchymal densities did not change significantly but the morphology of the calcifications was increasingly distorted as the compression rate increased leading to lower estimates of the suspiciousness of the cluster and higher uncertainties in the diagnosis. The uncompressed and two sets of compressed images were also processed by a wavelet method to extract the calcifications. Despite the fact that the segmentation algorithm generated several false-positive signals in highly compressed images, all true clusters were successfully segmented indicating that the compression process preserved the features of interest. Our preliminary results indicated that wavelets could be used to achieve high compression rates of mammographic images without losing small details such as microcalcification clusters as well as detect the calcifications from either the uncompressed or compressed reconstructed data. Further research and application of multiresolution analysis to digital mammography is continuing.