Free-response Operating Characteristic Research Articles

Assessing deep learning reconstruction for faster prostate MRI: visual vs. diagnostic performance metrics.

Deep learning (DL) MRI reconstruction enables fast scan acquisition with good visual quality, but the diagnostic impact is often not assessed because of large reader study requirements. This study used existing diagnostic DL to assess the diagnostic quality of reconstructed images. A retrospective multisite study of 1535 patients assessed biparametric prostate MRI between 2016 and 2020. Likely clinically significant prostate cancer (csPCa) lesions (PI-RADS 4) were delineated by expert radiologists. T2-weighted scans were retrospectively undersampled, simulating accelerated protocols. DL reconstruction (DLRecon) and diagnostic DL detection (DLDetect) were developed. The effect on the partial area under (pAUC), the Free-Response Operating Characteristic (FROC) curve, and the structural similarity (SSIM) were compared as metrics for diagnostic and visual quality, respectively. DLDetect was validated with a reader concordance analysis. Statistical analysis included Wilcoxon, permutation, and Cohen's kappa tests for visual quality, diagnostic performance, and reader concordance. DLRecon improved visual quality at 4- and 8-fold (R4, R8) subsampling rates, with SSIM (range: -1 to 1) improved to 0.78 ± 0.02 (p < 0.001) and 0.67 ± 0.03 (p < 0.001) from 0.68 ± 0.03 and 0.51 ± 0.03, respectively. However, diagnostic performance at R4 showed a pAUC FROC of 1.33 (CI 1.28-1.39) for DL and 1.29 (CI 1.23-1.35) for naive reconstructions, both significantly lower than fully sampled pAUC of 1.58 (DL: p = 0.024, naïve: p = 0.02). Similar trends were noted for R8. DL reconstruction produces visually appealing images but may reduce diagnostic accuracy. Incorporating diagnostic AI into the assessment framework offers a clinically relevant metric essential for adopting reconstruction models into clinical practice. In clinical settings, caution is warranted when using DL reconstruction for MRI scans. While it recovered visual quality, it failed to match the prostate cancer detection rates observed in scans not subjected to acceleration and DL reconstruction.

Application of a validated prostate MRI deep learning system to independent same-vendor multi-institutional data: demonstration of transferability

ObjectivesTo evaluate a fully automatic deep learning system to detect and segment clinically significant prostate cancer (csPCa) on same-vendor prostate MRI from two different institutions not contributing to training of the system.Materials and methodsIn this retrospective study, a previously bi-institutionally validated deep learning system (UNETM) was applied to bi-parametric prostate MRI data from one external institution (A), a PI-RADS distribution-matched internal cohort (B), and a csPCa stratified subset of single-institution external public challenge data (C). csPCa was defined as ISUP Grade Group ≥ 2 determined from combined targeted and extended systematic MRI/transrectal US-fusion biopsy. Performance of UNETM was evaluated by comparing ROC AUC and specificity at typical PI-RADS sensitivity levels. Lesion-level analysis between UNETM segmentations and radiologist-delineated segmentations was performed using Dice coefficient, free-response operating characteristic (FROC), and weighted alternative (waFROC). The influence of using different diffusion sequences was analyzed in cohort A.ResultsIn 250/250/140 exams in cohorts A/B/C, differences in ROC AUC were insignificant with 0.80 (95% CI: 0.74–0.85)/0.87 (95% CI: 0.83–0.92)/0.82 (95% CI: 0.75–0.89). At sensitivities of 95% and 90%, UNETM achieved specificity of 30%/50% in A, 44%/71% in B, and 43%/49% in C, respectively. Dice coefficient of UNETM and radiologist-delineated lesions was 0.36 in A and 0.49 in B. The waFROC AUC was 0.67 (95% CI: 0.60–0.83) in A and 0.7 (95% CI: 0.64–0.78) in B. UNETM performed marginally better on readout-segmented than on single-shot echo-planar-imaging.ConclusionFor same-vendor examinations, deep learning provided comparable discrimination of csPCa and non-csPCa lesions and examinations between local and two independent external data sets, demonstrating the applicability of the system to institutions not participating in model training.Clinical relevance statementA previously bi-institutionally validated fully automatic deep learning system maintained acceptable exam-level diagnostic performance in two independent external data sets, indicating the potential of deploying AI models without retraining or fine-tuning, and corroborating evidence that AI models extract a substantial amount of transferable domain knowledge about MRI-based prostate cancer assessment.Key Points• A previously bi-institutionally validated fully automatic deep learning system maintained acceptable exam-level diagnostic performance in two independent external data sets.• Lesion detection performance and segmentation congruence was similar on the institutional and an external data set, as measured by the weighted alternative FROC AUC and Dice coefficient.• Although the system generalized to two external institutions without re-training, achieving expected sensitivity and specificity levels using the deep learning system requires probability thresholds to be adjusted, underlining the importance of institution-specific calibration and quality control.

Artificial Intelligence Assisted Computational Tomographic Detection of Lung Nodules for Prognostic Cancer Examination: A Large-Scale Clinical Trial.

Low-dose computed tomography (LDCT) has emerged as a standard method for detecting early-stage lung cancer. However, the tedious computer tomography (CT) slide reading, patient-by-patient check, and lack of standard criteria to determine the vague but possible nodule leads to variable outcomes of CT slide interpretation. To determine the artificial intelligence (AI)-assisted CT examination, AI algorithm-assisted CT screening was embedded in the hospital picture archiving and communication system, and a 200 person-scaled clinical trial was conducted at two medical centers. With AI algorithm-assisted CT screening, the sensitivity of detecting nodules sized 4-5 mm, 6~10 mm, 11~20 mm, and &gt;20 mm increased by 41%, 11.2%, 10.3%, and 18.7%, respectively. Remarkably, the overall sensitivity of detecting varied nodules increased by 20.7% from 67.7% to 88.4%. Furthermore, the sensitivity increased by 18.5% from 72.5% to 91% for detecting ground glass nodules (GGN), which is challenging for radiologists and physicians. The free-response operating characteristic (FROC) AI score was ≥0.4, and the AI algorithm standalone CT screening sensitivity reached &gt;95% with an area under the localization receiver operating characteristic curve (LROC-AUC) of &gt;0.88. Our study demonstrates that AI algorithm-embedded CT screening significantly ameliorates tedious LDCT practices for doctors.

Locally adaptive decision in detection of clustered microcalcifications in mammograms

In computer-aided detection or diagnosis of clustered microcalcifications (MCs) in mammograms, the performance often suffers from not only the presence of false positives (FPs) among the detected individual MCs but also large variability in detection accuracy among different cases. To address this issue, we investigate a locally adaptive decision scheme in MC detection by exploiting the noise characteristics in a lesion area. Instead of developing a new MC detector, we propose a decision scheme on how to best decide whether a detected object is an MC or not in the detector output. We formulate the individual MCs as statistical outliers compared to the many noisy detections in a lesion area so as to account for the local image characteristics. To identify the MCs, we first consider a parametric method for outlier detection, the Mahalanobis distance detector, which is based on a multi-dimensional Gaussian distribution on the noisy detections. We also consider a non-parametric method which is based on a stochastic neighbor graph model of the detected objects. We demonstrated the proposed decision approach with two existing MC detectors on a set of 188 full-field digital mammograms (95 cases). The results, evaluated using free response operating characteristic (FROC) analysis, showed a significant improvement in detection accuracy by the proposed outlier decision approach over traditional thresholding (the partial area under the FROC curve increased from 3.95 to 4.25, p-value <10−4). There was also a reduction in case-to-case variability in detected FPs at a given sensitivity level. The proposed adaptive decision approach could not only reduce the number of FPs in detected MCs but also improve case-to-case consistency in detection.

Computer-aided detection of breast lesions in DCE-MRI using region growing based on fuzzy C-means clustering and vesselness filter

A computer-aided detection (CAD) system is introduced in this paper for detection of breast lesions in dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI). The proposed CAD system firstly compensates motion artifacts and segments the breast region. Then, the potential lesion voxels are detected and used as the initial seed points for the seeded region-growing algorithm. A new and robust region-growing algorithm incorporating with Fuzzy C-means (FCM) clustering and vesselness filter is proposed to segment any potential lesion regions. Subsequently, the false positive detections are reduced by applying a discrimination step. This is based on 3D morphological characteristics of the potential lesion regions and kinetic features which are fed to the support vector machine (SVM) classifier. The performance of the proposed CAD system is evaluated using the free-response operating characteristic (FROC) curve. We introduce our collected dataset that includes 76 DCE-MRI studies, 63 malignant and 107 benign lesions. The prepared dataset has been used to verify the accuracy of the proposed CAD system. At 5.29 false positives per case, the CAD system accurately detects 94% of the breast lesions.

Breast cancer detection in automated 3D breast ultrasound using iso-contours and cascaded RUSBoosts

Automated 3D breast ultrasound (ABUS) is a new popular modality as an adjunct to mammography for detecting cancers in women with dense breasts. In this paper, a multi-stage computer aided detection system is proposed to detect cancers in ABUS images. In the first step, an efficient despeckling method called OBNLM is applied on the images to reduce speckle noise. Afterwards, a new algorithm based on isocontours is applied to detect initial candidates as the boundary of masses is hypo echoic. To reduce false generated isocontours, features such as hypoechoicity, roundness, area and contour strength are used. Consequently, the resulted candidates are further processed by a cascade classifier whose base classifiers are Random Under-Sampling Boosting (RUSBoost) that are introduced to deal with imbalanced datasets. Each base classifier is trained on a group of features like Gabor, LBP, GLCM and other features. Performance of the proposed system was evaluated using 104 volumes from 74 patients, including 112 malignant lesions. According to Free Response Operating Characteristic (FROC) analysis, the proposed system achieved the region-based sensitivity and case-based sensitivity of 68% and 76% at one false positive per image.

Extended PCJO for the Detection-Localization of Hypersignals and Hyposignals in CT Images

The anthropomorphic model observer (MO) plays an important role in the assessment and optimization of medical imaging systems. The MO is a task-based approach; while the abnormality can appear as a hypersignal or a hyposignal for different imaging modalities, sequences, or organs, no MO has been proposed for the hyposignals detection-localization task in the literature. To improve the clinical relevance of the existing MOs, we propose an anthropomorphic MO that can also deal with hyposignals in this paper. In a previous study, we reported a perceptually relevant channelized joint observer (PCJO) for detecting and localizing multiple signals with unknown amplitude, orientation, size, and location. Here, we extend it mathematically to hyposignals task. A free-response study (close to the real-diagnostic procedure) for both hypersignals and hyposignals in cerebral and abdominal CT images was conducted with four radiologists. The equally weighted alternative free-response operating characteristic was used as the figure of merit. Statistical analyses show that the extended PCJO approaches the experts’ performances with no significant difference in the studied tasks. The results demonstrate that the extended PCJO is an alternative to replace radiologists for the evaluation and comparison of different medical image processing algorithms. The PCJO has been originally proposed on magnetic resonance imaging but tested on computerized tomography (CT) here; the coherent results show that the PCJO can be generalized to another modality-CT. We also provide in this paper, the reference values of all the parameters in the PCJO to facilitate its future application on magnetic resonance (MR) or CT images.

Performance analysis of cone detection algorithms

Many algorithms have been proposed to help clinicians evaluate cone density and spacing, as these may be related to the onset of retinal diseases. However, there has been no rigorous comparison of the performance of these algorithms. In addition, the performance of such algorithms is typically determined by comparison with human observers. Here we propose a technique to simulate realistic images of the cone mosaic. We use the simulated images to test the performance of three popular cone detection algorithms, and we introduce an algorithm which is used by astronomers to detect stars in astronomical images. We use Free Response Operating Characteristic (FROC) curves to evaluate and compare the performance of the four algorithms. This allows us to optimize the performance of each algorithm. We observe that performance is significantly enhanced by up-sampling the images. We investigate the effect of noise and image quality on cone mosaic parameters estimated using the different algorithms, finding that the estimated regularity is the most sensitive parameter.

Automated localization of breast cancer in DCE-MRI.

Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is increasingly being used for the detection and diagnosis of breast cancer. Compared to mammography, DCE-MRI provides higher sensitivity, however its specificity is variable. Moreover, DCE-MRI data analysis is time consuming and depends on reader expertise. The aim of this work is to propose a novel automated breast cancer localization system for DCE-MRI. Such a system can be used to support radiologists in DCE-MRI analysis by marking suspicious areas. The proposed method initially corrects for motion artifacts and segments the breast. Subsequently, blob and relative enhancement voxel features are used to locate lesion candidates. Finally, a malignancy score for each lesion candidate is obtained using region-based morphological and kinetic features computed on the segmented lesion candidate. We performed experiments to compare the use of different classifiers in the region classification stage and to study the effect of motion correction in the presented system. The performance of the algorithm was assessed using free-response operating characteristic (FROC) analysis. For this purpose, a dataset of 209 DCE-MRI studies was collected. It is composed of 95 DCE-MRI studies with 105 breast cancers (55 mass-like and 50 non-mass-like malignant lesions) and 114 DCE-MRI studies from women participating in a screening program which were diagnosed to be normal. At 4 false positives per normal case, 89% of the breast cancers (91% and 86% for mass-like and non-mass-like malignant lesions, respectively) were correctly detected.

A new automatic mass detection method for breast cancer with false positive reduction

Mass localization is a crucial problem in computer-aided detection (CAD) system for the diagnosis of suspicious regions in mammograms. In this paper, a new automatic mass detection method for breast cancer in mammographic images is proposed. Firstly, suspicious regions are located with an adaptive region growing method, named multiple concentric layers (MCL) approach. Prior knowledge is utilized by tuning parameters with training data set during the MCL step. Then, the initial regions are further refined with narrow band based active contour (NBAC), which can improve the segmentation accuracy of masses. Texture features and geometry features are extracted from the regions of interest (ROI) containing the segmented suspicious regions and the boundaries of the segmentation. The texture features are computed from gray level co-occurrence matrix (GLCM) and completed local binary pattern (CLBP). Finally, the ROIs are classified by means of support vector machine (SVM), with supervision provided by the radiologist׳s diagnosis. To deal with the imbalance problem regarding the number of non-masses and masses, supersampling and downsampling are incorporated. The method was evaluated on a dataset with 429 craniocaudal (CC) view images, containing 504 masses. Among them, 219 images containing 260 masses are used to optimize the parameters during MCL step, and are used to train SVM. The remaining 210 images (with 244 masses) are used to test the performance. Masses are detected with 82.4% sensitivity with 5.3 false positives per image (FPsI) with MCL, and after active contour refinement, feature analysis and classification, it obtained 1.48 FPsI at the sensitivity 78.2%. Testing on 164 normal mammographic images showed 5.18 FPsI with MCL and 1.51 FPsI after classification. Experiments on mediolateral oblique (MLO) images have also been performed, the proposed method achieved a sensitivity 75.6% at 1.38 FPsI. The method is also analyzed with free response operating characteristic (FROC) and compared with previous methods. Overall, the proposed method is a promising approach to achieve low FPsI while maintaining a high sensitivity.

Computer-aided diagnosis of breast microcalcifications based on dual-tree complex wavelet transform

BackgroundDigital mammography is the most reliable imaging modality for breast carcinoma diagnosis and breast micro-calcifications is regarded as one of the most important signs on imaging diagnosis. In this paper, a computer-aided diagnosis (CAD) system is presented for breast micro-calcifications based on dual-tree complex wavelet transform (DT-CWT) to facilitate radiologists like double reading.MethodsFirstly, 25 abnormal ROIs were extracted according to the center and diameter of the lesions manually and 25 normal ROIs were selected randomly. Then micro-calcifications were segmented by combining space and frequency domain techniques. We extracted three texture features based on wavelet (Haar, DB4, DT-CWT) transform. Totally 14 descriptors were introduced to define the characteristics of the suspicious micro-calcifications. Principal Component Analysis (PCA) was used to transform these descriptors to a compact and efficient vector expression. Support Vector Machine (SVM) classifier was used to classify potential micro-calcifications. Finally, we used the receiver operating characteristic (ROC) curve and free-response operating characteristic (FROC) curve to evaluate the performance of the CAD system.ResultsThe results of SVM classifications based on different wavelets shows DT-CWT has a better performance. Compared with other results, DT-CWT method achieved an accuracy of 96% and 100% for the classification of normal and abnormal ROIs, and the classification of benign and malignant micro-calcifications respectively. In FROC analysis, our CAD system for clinical dataset detection achieved a sensitivity of 83.5% at a false positive per image of 1.85.ConclusionsCompared with general wavelets, DT-CWT could describe the features more effectively, and our CAD system had a competitive performance.

Effect of image quality on calcification detection in digital mammography

This study aims to investigate if microcalcification detection varies significantly when mammographic images are acquired using different image qualities, including: different detectors, dose levels, and different image processing algorithms. An additional aim was to determine how the standard European method of measuring image quality using threshold gold thickness measured with a CDMAM phantom and the associated limits in current EU guidelines relate to calcification detection. One hundred and sixty two normal breast images were acquired on an amorphous selenium direct digital (DR) system. Microcalcification clusters extracted from magnified images of slices of mastectomies were electronically inserted into half of the images. The calcification clusters had a subtle appearance. All images were adjusted using a validated mathematical method to simulate the appearance of images from a computed radiography (CR) imaging system at the same dose, from both systems at half this dose, and from the DR system at quarter this dose. The original 162 images were processed with both Hologic and Agfa (Musica-2) image processing. All other image qualities were processed with Agfa (Musica-2) image processing only. Seven experienced observers marked and rated any identified suspicious regions. Free response operating characteristic (FROC) and ROC analyses were performed on the data. The lesion sensitivity at a nonlesion localization fraction (NLF) of 0.1 was also calculated. Images of the CDMAM mammographic test phantom were acquired using the automatic setting on the DR system. These images were modified to the additional image qualities used in the observer study. The images were analyzed using automated software. In order to assess the relationship between threshold gold thickness and calcification detection a power law was fitted to the data. There was a significant reduction in calcification detection using CR compared with DR: the alternative FROC (AFROC) area decreased from 0.84 to 0.63 and the ROC area decreased from 0.91 to 0.79 (p < 0.0001). This corresponded to a 30% drop in lesion sensitivity at a NLF equal to 0.1. Detection was also sensitive to the dose used. There was no significant difference in detection between the two image processing algorithms used (p > 0.05). It was additionally found that lower threshold gold thickness from CDMAM analysis implied better cluster detection. The measured threshold gold thickness passed the acceptable limit set in the EU standards for all image qualities except half dose CR. However, calcification detection varied significantly between image qualities. This suggests that the current EU guidelines may need revising. Microcalcification detection was found to be sensitive to detector and dose used. Standard measurements of image quality were a good predictor of microcalcification cluster detection.

Nonparametric signal detectability evaluation using an exponential transformation of the FROC curve

To develop an efficient nonparametric method for evaluation the detectability of signals at unknown locations in images, as a mean for image quality assessment. We use the free-response methodology that allows the image observer to mark and score all locations found as suspicious in an image, summarizing these results in a free-response operating characteristic (FROC) curve. However, unlike the relative (or receiver) operating characteristic (ROC), or the localization ROC (LROC), the FROC curve has an undefined, theoretically infinite, right side limit. Therefore area under the FROC curves cannot be directly used as an overall performance index, as the area under the curve is for ROC or LROC. We circumvent this drawback by using a transformation of the abscissa that leads to a finite integration range. By applying an exponential transformation we derive a nonparametric estimator for such a metric, and we study its properties by deriving analytical expressions for the mean and standard deviation in conditions of scores independence. A comparative study with other related nonparametric estimators for ROC, LROC, and alternative FROC (AFROC) method is presented. The new nonparametric estimator has sensitivity and scalability properties that make it particularly advantageous for signal detectability evaluation in phantom experiments using model observers.

Detection of microcalcifications in digital mammograms using combined model-based and statistical textural features

We investigated the performance of clustered microcalcifications (MCs) recognition in digital mammograms by using combined model-based and statistical textural features. Twenty mammograms containing 25 areas of MCs from the MIAS MiniMammogram database were used to test the performance of our method. In the first stage, a wavelet filter and two thresholds were used to detect suspicious MCs from the mammograms. In the second stage, textural features based on Markov random field (MRF) and fractal models together with statistical textural features based on the surrounding region-dependence method (SRDM) were extracted from the neighborhood of the suspicious MCs and were classified by a three-layer BPNN. The free-response operating characteristic (FROC) curve was used to evaluate the performance of classification and compare our results with that presented in the literature from four other studies. The results of the experiments suggest that the combined model-based and statistical textural features are suitable for characterizing microcalcifications and capable of supporting a reliable and effective MCs detection. In particular, a true positive rate of about 94% is achieved at the rate of 1.0 false positive per image, or the false positives per image can be reduced to 0.65 FPs/image at the rate of true positive of about 90%.

Perceptually Based FROC Analysis 1

Analysis of reading data when cases have multiple targets and/or the reader is required to localize targets is difficult. One approach to this free-response operating characteristic (FROC) problem is for images to be segmented (eg, with quadrants) by the investigator and a segment-level analysis be conducted with the case as a nesting factor. In this report, we introduce an alternative method that uses the visual scan path of the reader to segment the image. We evaluate the new method by applying it to data from a mammography reading experiment. The gaze scan path of one radiologist was recorded as she scanned 40 mammograms for masses and microcalcifications. The observer is an experienced mammographer and was not one of the authors. In addition, the reader provided a rating indicating the degree of suspicion for any suspected targets she identified and localized. We then established "perceptual regions" by using a clustering algorithm on the visual fixations. We combined ratings given to specific locations indicated by the reader with the segmentation from the visual scan to generate a series of ratings classified for whether the perceptually based region associated with the rating contained or did not contain a known target. We analyzed data generated by our method from all 40 cases by using the conventional maximum-likelihood method based on the binormal model. Finally, we tested goodness-of-fit of the binormal model to the data by using chi-square. Maximum-likelihood estimation led to a model that did not fit the data (P < .001). However, examination of the observed and expected counts suggests that the binormal assumption does not hold for segments that contain targets and a bimodal distribution model might be preferred. Our new method provides an alternative approach to analysis of the FROC experiment. It needs to be developed further. Specifically, we propose that a mixture model extension of the binormal model be developed for ratings data arising from perceptually based FROC experiments. A disadvantage to our method is the requirement to record the scan path of the reader. However, we believe that adding such information to receiver operating characteristic (ROC) curve analysis will pay off when appropriate statistical models have been identified because we believe our data support our hypothesis that the perceptual scanning of images by humans deconvolves interpretation correlation. If true, this hypothesis implies that conventional statistical methods for ROC analysis based on independent data can be applied to the analysis of FROC data after conditioning on the scan path of the observer.

Interval change analysis to improve computer aided detection in mammography

We are developing computer aided diagnosis (CAD) techniques to study interval changes between two consecutive mammographic screening rounds. We have previously developed methods for the detection of malignant masses based on features extracted from single mammographic views. The goal of the present work was to improve our detection method by including temporal information in the CAD program. Toward this goal, we have developed a regional registration technique. This technique links a suspicious location on the current mammogram with a corresponding location on the prior mammogram. The novelty of our method is that the search for correspondence is done in feature space. This has the advantage that very small lesions and architectural distortions may be found as well. Following the linking process several features are calculated for the current and prior region. Temporal features are obtained by combining the feature values from both regions. We evaluated the detection performance with and without the use of temporal features on a data set containing 2873 temporal film pairs from 938 patients. There were 589 cases in which the current mammogram contained exactly one malignant mass. Cross validation was used to partition the data set into a train set and a test set. The train set was used for feature selection and classifier training, the test set for classifier evaluation. FROC (free response operating characteristic) analysis showed an improvement in detection performance with the use of temporal features.

Assessing target search performance: the free-response operator characteristic model

The receiver operating characteristic (ROC) curve is a standard method for quantifying the performance of a detection task where the signal of interest is embedded in noise. ROC analysis has been applied to a variety of signal detection problems including medical imaging, acoustics, and automated target detection (ATD). The free response operating characteristic (FROC) curve generalizes the ROC model for search problems. The FROC model is appropriate when multiple detections are possible and the number of false alarms is unconstrained. The shape of the FROC curve depends on the underlying probability distributions for the signal and noise. The general FROC model is presented and parameter estimation is discussed. An example illustrates the approach.

Artificial neural networks for single photon emission computed tomography. A study of cold lesion detection and localization.

An artificial neural network was developed for cold lesion detection and localization in single photon emission computed tomography (SPECT) images. The network was trained for several noise levels and lesion sizes to identify lesions located in the center of small image neighborhoods. When scrolled across an image the trained network was able to identify cold abnormalities. The diagnostic performance of the technique was evaluated at two noise levels (50,000 and 100,000 counts/slice) and for two lesion sizes (radius: 1.0 cm and 1.5 cm) using the free-response operating characteristic (FROC) analysis. Furthermore, the same network was tested on a situation it was not trained on (80,000 counts/slice and a different reconstruction filter). The neural network showed high sensitivity and small false-positive rates per image for all test situations. These results suggest that neural networks are promising tools for computer-aided clinical diagnosis in SPECT: