ABSTRACT
Biomedical applications of deep learning algorithms rely on large expert annotated data sets. The classification of bone marrow (BM) cell cytomorphology, an important cornerstone of hematological diagnosis, is still done manually thousands of times every day because of a lack of data sets and trained models. We applied convolutional neural networks (CNNs) to a large data set of 171 374 microscopic cytological images taken from BM smears from 945 patients diagnosed with a variety of hematological diseases. The data set is the largest expert-annotated pool of BM cytology images available in the literature. It allows us to train high-quality classifiers of leukocyte cytomorphology that identify a wide range of diagnostically relevant cell species with high precision and recall. Our CNNs outcompete previous feature-based approaches and provide a proof-of-concept for the classification problem of single BM cells. This study is a step toward automated evaluation of BM cell morphology using state-of-the-art image-classification algorithms. The underlying data set represents an educational resource, as well as a reference for future artificial intelligence-based approaches to BM cytomorphology.
Subject(s)
Bone Marrow Cells/pathology , Hematologic Diseases/diagnosis , Neural Networks, Computer , Bone Marrow Cells/cytology , Cell Differentiation , Hematologic Diseases/pathology , Humans , Image Processing, Computer-Assisted/methods , Microscopy/methodsABSTRACT
INTRODUCTION: Although mammographic density is an established risk factor for breast cancer, its use is limited in clinical practice because of a lack of automated and standardized measurement methods. The aims of this study were to evaluate a variety of automated texture features in mammograms as risk factors for breast cancer and to compare them with the percentage mammographic density (PMD) by using a case-control study design. METHODS: A case-control study including 864 cases and 418 controls was analyzed automatically. Four hundred seventy features were explored as possible risk factors for breast cancer. These included statistical features, moment-based features, spectral-energy features, and form-based features. An elaborate variable selection process using logistic regression analyses was performed to identify those features that were associated with case-control status. In addition, PMD was assessed and included in the regression model. RESULTS: Of the 470 image-analysis features explored, 46 remained in the final logistic regression model. An area under the curve of 0.79, with an odds ratio per standard deviation change of 2.88 (95% CI, 2.28 to 3.65), was obtained with validation data. Adding the PMD did not improve the final model. CONCLUSIONS: Using texture features to predict the risk of breast cancer appears feasible. PMD did not show any additional value in this study. With regard to the features assessed, most of the analysis tools appeared to reflect mammographic density, although some features did not correlate with PMD. It remains to be investigated in larger case-control studies whether these features can contribute to increased prediction accuracy.
Subject(s)
Breast Neoplasms/diagnostic imaging , Mammography/methods , Aged , Area Under Curve , Breast Neoplasms/pathology , Case-Control Studies , Female , Humans , Logistic Models , Middle Aged , Odds Ratio , Risk FactorsABSTRACT
PURPOSE: To compare diagnostic performance and interobserver variability in a group of 36 examiners, with four different levels of experience. METHODS: Nine junior trainees, eight level I senior trainees, 11 level II senior gynecologists, and eight level III expert sonologists classified 105 ultrasound images of adnexal masses into three subgroups of ovarian lesions (malignancies, functional cysts, and dermoid cysts). RESULTS: The level III sonologists obtained the best diagnostic results together with the lowest interobserver variability (κ = 0.70, SD = 0.04). They achieved significantly better results in comparison with the junior trainees and also the senior trainees (κ = 0.51, SD = 0.12, p < 0.001; and κ = 0.51, SD = 0.09, p < 0.001). Differences between level III sonologists and the group of level II observers did not reach statistical significance (κ = 0.65, SD = 0.09, p = 0.70). There were no significant differences between senior and junior trainees (p = 1.0) and both groups achieved a significantly poorer diagnostic performance in comparison with the level II observers (p < 0.01 and p < 0.01). For all observers, the largest differences were seen for classifying malignancies, the best results for classifying functional cysts, and the poorest for evaluating dermoid cysts. CONCLUSIONS: Diagnostic performance of pattern recognition significantly improves with an increasing level of experience, emphasizing the importance of standardized ultrasound training programs with supervision by experts.
Subject(s)
Ovarian Diseases/pathology , Ovary/pathology , Pattern Recognition, Visual , Adolescent , Adult , Aged , Aged, 80 and over , Female , Humans , Middle Aged , Observer Variation , Organ Size , Ovarian Diseases/diagnostic imaging , Ovary/diagnostic imaging , Ultrasonography , Young AdultABSTRACT
PURPOSE: Diagnosis of cervical intraepithelial neoplasia (CIN) is currently based on the histological result of an aiming biopsy. This preliminary study investigated whether diagnostics for CIN can potentially be improved using semiautomatic colposcopic image analysis. METHODS: 198 women with unremarkable or abnormal smears underwent colposcopy examinations. 375 regions of interest (ROIs) were manually marked on digital screen shots of the streaming documentation, which we provided during our colposcopic examinations (39 normal findings, 41 CIN I, and 118 CIN II-III). These ROIs were classified into two groups (211 regions with normal findings and CIN I, and 164 regions with CIN II-III). We developed a prototypical computer-assisted diagnostic (CAD) device based on image-processing methods to automatically characterize the color, texture, and granulation of the ROIs. RESULTS: Using n-fold cross-validation, the CAD system achieved a maximum diagnostic accuracy of 80% (sensitivity 85% and specificity 75%) corresponding to a correct assignment of abnormal or unremarkable findings. CONCLUSIONS: The CAD system may be able to play a supportive role in the further diagnosis of CIN, potentially paving the way for new and enhanced developments in colposcopy-based diagnosis.
Subject(s)
Colposcopy/methods , Diagnosis, Computer-Assisted/methods , Uterine Cervical Dysplasia/diagnosis , Uterine Cervical Neoplasms/diagnosis , Adult , Aged , Cervix Uteri/pathology , Female , Humans , Image Processing, Computer-Assisted , Middle Aged , Reproducibility of Results , Sensitivity and Specificity , Vaginal SmearsABSTRACT
This paper presents a new spectral approach to color correction for medical image analysis applications. Linear estimation with regularization by a constrained principal eigenvector method is used for calibration of the camera system and estimation of the illumination spectrum while spectral surface reflectivities are determined by Wiener inverse estimation. Nonlinear devices are handled by piecewise linear interpolation and any linear color preprocessing inside the camera is explicitly modeled. All measurement and estimation processes are combined into a spectral calibration framework for practical application in computer-assisted image analysis. The novelty of our approach lies in the generalization of the image formation model allowing for linear preprocessing inside the camera system. Such transforms would lead to erroneous results with positivity constraint based algorithms or a monochromator based measurement. We provide experimental results from a comprehensive set of reference measurements acquired with a video endoscopy system for gastroscopic application.
Subject(s)
Algorithms , Artifacts , Colorimetry/methods , Gastroscopy/methods , Image Enhancement/methods , Image Interpretation, Computer-Assisted/methods , Video Recording/methods , Colorimetry/instrumentation , Gastroscopes , Phantoms, Imaging , Reproducibility of Results , Sensitivity and Specificity , Video Recording/instrumentationABSTRACT
Modern techniques for medical diagnostics and therapy in keyhole surgery scenarios as well as technical inspection make use of flexible endoscopes. Their characteristic bendable image conductor consists of a very limited number of coated fibers, which leads to so-called comb structure. This effect has a negative impact on further image processing steps such as feature tracking because these overlaid image structures are wrongly detected as image features. With respect to these tasks, we propose an automatic approach to generate optimal spectral filter masks for enhancement of fiberscopic images. We apply the Nyquist-Shannon sampling theorem to the spectrum of fiberscopically acquired images to obtain parameters for optimal filter mask calculation. This can be done automatically and independently of scale and resolution of the image conductor as well as type and resolution of the image sensor. We designed and verified simple rotation invariant masks as well as star-shaped rotation variant masks that contain information about orientation between the fiberscope and sensor. A subjective survey among experts between different modes of filtering certified the best results to the adapted star-shaped mask for high-quality glass fiberscopes. We furthermore define an objective metric to evaluate the results of different filter approaches, which verifies the results of the subjective survey. The proposed approach enables the automated reduction of fiberscopic comb structure. It is adaptive to arbitrary endoscope and sensor combinations. The results give the prospect of a large field of possible applications to reduce fiberscopic structure both for visual optimization in clinical environments and for further digital imaging tasks.
Subject(s)
Artificial Intelligence , Endoscopes , Endoscopy/methods , Fiber Optic Technology/instrumentation , Image Enhancement/methods , Image Interpretation, Computer-Assisted/methods , Pattern Recognition, Automated/methods , Algorithms , Equipment Design , Equipment Failure Analysis , Fiber Optic Technology/methods , Imaging, Three-Dimensional/methods , Information Storage and Retrieval/methods , Phantoms, Imaging , Reproducibility of Results , Sensitivity and Specificity , Signal Processing, Computer-AssistedABSTRACT
PURPOSE: Inhomogeneous illumination often causes significant shading and vignetting effects in images captured by an endoscope. Most of the established shading correction methods are designed for gray-level images. Only few papers have been published about how to compensate for shading in color images. For endoscopic images with a distinct red coloring, these methods tend to produce color artifacts. METHOD: A color shading correction algorithm for endoscopic images is proposed. Principal component analysis is used to calculate an appropriate estimate of the shading effect so that a one-channel shading correction can be applied without producing undesired artifacts. RESULTS: The proposed method is compared to established YUV and HSV color-conversion-based approaches. It produces superior results both on simulated and on real endoscopic images. Example images of using the proposed shading correction for endoscopic image mosaicking are presented. CONCLUSION: A new method for shading correction is presented which is tailored to images with distinct coloring. It is beneficial for the visual impression and further image analysis tasks.
Subject(s)
Algorithms , Phantoms, Imaging , Ureteroscopy/methods , Urinary Bladder/diagnostic imaging , Color , Humans , Image Interpretation, Computer-Assisted , Lighting , Pattern Recognition, Automated , Radiography , Reproducibility of Results , Subtraction TechniqueABSTRACT
BACKGROUND: A prototype system for computer-assisted colposcopic diagnosis (CAD) currently achieves a high level of accuracy of 80% (sensitivity 85%, specificity 75%) for the automatic assessment of colposcopic images. This pilot study investigated whether this type of CAD system is, in principle, capable of influencing the quality of the examiner's assessment. MATERIALS AND METHODS: In this observer study, 24 digitized colposcopic images from patients attending a dysplasia clinic were assessed by 90 participants. All participants had attended a colposcopy training workshop so that they acquired the same basic information and skills. RESULTS: Wide variation was seen among the non-experts, in contrast to the experts. An overall improvement in diagnostic accuracy was noted when the CAD system was used (non-experts: sensitivity 78%, specificity 70%; experts: sensitivity 74%, specificity 70%). CONCLUSION: The CAD system may serve as an aid in the further diagnosis of cervical intraepithelial neoplasia, and has the potential to improve the diagnostic process.
Subject(s)
Colposcopy/methods , Diagnosis, Computer-Assisted/methods , Uterine Cervical Dysplasia/diagnosis , Uterine Cervical Neoplasms/diagnosis , Colposcopy/education , Colposcopy/standards , Diagnosis, Computer-Assisted/standards , Female , Humans , Observer Variation , Pilot Projects , Sensitivity and Specificity , Uterine Cervical Neoplasms/pathology , Uterine Cervical Dysplasia/pathologyABSTRACT
Flexible endoscopes based on fiber bundles are still widely used despite the recent success of so-called tipchip endoscopes. This is partly due to the costs and that for extremely thin diameters (below 3 mm) there are still only fiberscopes available. Due to the inevitable artifacts caused by the transition from the fiber bundles to the sensor chip, image and texture analysis algorithms are severely handicapped. Therefore, texture-based computer-assisted diagnosis (CAD) systems could not be used in such domains without image preprocessing. We describe a CAD system approach that includes an image filtering algorithm to remove the fiber image artifacts first and then applies conventional color texture algorithms that have been applied to other endoscopic disciplines in the past. The concept is evaluated on an image database with artificially rendered fiber artifacts so that ground truth information is available.