Fine-Tuning and training of densenet for histopathology image representation using TCGA diagnostic slides.

Feature vectors provided by pre-trained deep artificial neural networks have become a dominant source for image representation in recent literature. Their contribution to the performance of image analysis can be improved through fine-tuning. As an ultimate solution, one might even train a deep network from scratch with the domain-relevant images, a highly desirable option which is generally impeded in pathology by lack of labeled images and the computational expense. In this study, we propose a new network, namely KimiaNet, that employs the topology of the DenseNet with four dense blocks, fine-tuned and trained with histopathology images in different configurations. We used more than 240,000 image patches with 1000×1000 pixels acquired at 20× magnification through our proposed "high-cellularity mosaic" approach to enable the usage of weak labels of 7126 whole slide images of formalin-fixed paraffin-embedded human pathology samples publicly available through The Cancer Genome Atlas (TCGA) repository. We tested KimiaNet using three public datasets, namely TCGA, endometrial cancer images, and colorectal cancer images by evaluating the performance of search and classification when corresponding features of different networks are used for image representation. As well, we designed and trained multiple convolutional batch-normalized ReLU (CBR) networks. The results show that KimiaNet provides superior results compared to the original DenseNet and smaller CBR networks when used as feature extractor to represent histopathology images.

Recognizing Magnification Levels in Microscopic Snapshots.

Recent advances in digital imaging has transformed computer vision and machine learning to new tools for analyzing pathology images. This trend could automate some of the tasks in the diagnostic pathology and elevate the pathologist workload. The final step of any cancer diagnosis procedure is performed by the expert pathologist. These experts use microscopes with high level of optical magnification to observe minute characteristics of the tissue acquired through biopsy and fixed on glass slides. Switching between different magnifications, and finding the magnification level at which they identify the presence or absence of malignant tissues is important. As the majority of pathologists still use light microscopy, compared to digital scanners, in many instance a mounted camera on the microscope is used to capture snapshots from significant field- of-views. Repositories of such snapshots usually do not contain the magnification information. In this paper, we extract deep features of the images available on TCGA dataset with known magnification to train a classifier for magnification recognition. We compared the results with LBP, a well-known handcrafted feature extraction method. The proposed approach achieved a mean accuracy of 96% when a multi-layer perceptron was trained as a classifier.

Restricting motion effects in CT coronary angiography.

OBJECTIVE: Evaluation of coronary CT image blur using multi segment reconstruction algorithm. METHODS: Cardiac motion was simulated in a Catphan. CT coronary angiography was performed using 320 × 0.5 mm detector array and 275 ms gantry rotation. 1, 2 and 3 segment reconstruction algorithm, three heart rates (60, 80 and 100bpm), two peak displacements (4, 8 mm) and three cardiac phases (55, 35, 75%) were used. Wilcoxon test compared image blur from the different reconstruction algorithms. RESULTS: Image blur for 1, 2 and 3 segments in: 60 bpm, 75% R-R interval and 8 mm peak displacement: 0.714, 0.588, 0.571 mm (1.18, 0.6, 0.4 mm displacement) 80 bpm, 35% R-R interval and 8 mm peak displacement: 0.869, 0.606, 0.606 mm (1.57, 0.79,0.52 mm displacement) 100 bpm, 35% R-R interval and 4 mm peak displacement: 0.645, 0.588, 0.571 mm (0.98, 0.49, 0.33 mm displacement). The median image blur overall for 1 and 2 segments was 0.714 mm and 0.588 mm respectively (p < 0.0001). CONCLUSION: Two-segment reconstruction significantly reduces image blur. ADVANCES IN KNOWLEDGE: Multisegment reconstruction algorithms during CT coronary angiography are a useful method to reduce image blur, improve visualization of the coronary artery wall and help the early detection of the plaque.

Optimization of Computed Tomography Coronary Angiography for Improved Plaque Detection.

OBJECTIVE: The study aims to optimize visualization of the coronary wall during computed tomography coronary angiography. METHODS: A coronary plaque phantom was scanned on a wide-volume computed tomography scanner. Spatial resolution, contrast resolution, and vessel wall thickness were measured at different x-ray tube currents and voltages. RESULTS: Spatial resolution ranged from 0.385 to 0.625 mm and was significantly lower at higher currents. Contrast-to-noise ratio was significantly higher at higher currents. The most accurate wall thickness measurements were quantified at 300 and 400 mA for 80 and 100 kVp and 300 mA for 120 and 135 kVp. CONCLUSIONS: Lower spatial resolution at higher currents was due to added blur from increased focal spot size. Contrast-to-noise ratio was higher at higher currents owing to decreased quantum noise. Wall thickness was measured more accurately at intermediate currents with midrange contrast-to-noise ratio but optimal spatial resolution. For accurate coronary wall thickness measurement, contrast-to-noise ratio is compromised to achieve optimal spatial resolution.

Image quality and radiation dose of pulmonary CT angiography performed using 100 and 120 kVp.

OBJECTIVE: The objective of our study was to compare image quality and radiation dose of pulmonary CT angiography (CTA) performed in the same patient cohort using tube potentials of 100 and 120 kVp. MATERIALS AND METHODS: The study group for this retrospective study was 32 patients (22 women, 10 men) with a mean age of 57 years (age range, 28-83 years; body weight < 100 kg). Patients underwent pulmonary CTA studies performed using 120 and 100 kVp while other scanning parameters were kept constant. Two observers measured image signal and image noise, signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR), and SNR dose and CNR dose. Two additional observers performed qualitative image quality analysis using a 5-point grading scale (5 = excellent). RESULTS: The reduction in tube potential caused image signal to increase by 29% (p < 0.0001), image noise to increase by 68% (p < 0.0001), CNR dose to decrease by 0.8% (p = 0.91) and SNR to decrease by 24% (p = 0.0002) and CNR by 20% (p = 0.0019). Radiation dose (dose-length product) was decreased by 37% to 379.26 mGy × cm at 100 kVp from 604.46 mGy × cm at 120 kVp (p < 0.0001). The median pulmonary arteries image quality scores for observers 1 and 2, respectively, were as follows at 100 kVp: main, 5 and 5; lobar, 5 and 4.5; and segmental, 5 and 4. At 120 kVp, the median image quality scores for observers 1 and 2 were as follows: main, 5 and 5; lobar, 5 and 5; segmental, 4 and 4. A Wilcoxon test analysis indicated no significant difference in image quality between the studies (main, p = 0.59; lobar, p = 0.88; segmental, p = 0.79). CONCLUSION: Pulmonary CTA can be performed using a tube potential of 100 kVp in patients who weigh less than 100 kg (220 lb). Reducing the tube potential from 120 to 100 kVp results in a 37% reduction in radiation dose without a significant impact on diagnostic image quality.

The influence of chest wall tissue composition in determining image noise during cardiac CT.

OBJECTIVE: The purpose of this article is to determine the influence of chest wall composition on image quality in cardiac CT. MATERIALS AND METHODS: A retrospective study of 100 consecutive patients referred for CT coronary artery calcium assessment was performed. Image noise (Hounsfield units) was measured by prescribing a region of interest in the descending thoracic aorta. Image noise was correlated with conventional patient biometric parameters, including body weight, body mass index (BMI), and anteroposterior and lateral thoracic diameters, and with novel patient biometric parameters, including total chest wall soft tissue, chest wall fat, and chest wall muscle and bone. The linear correlation coefficient was used to indicate the strength of the association. RESULTS: A strong correlation was noted between BMI and image noise in men (r = 0.66), but the strongest relationships were observed in larger women (BMI ≥ 25), who had more chest wall fat than muscle and very strong correlations between image noise, chest wall fat (r = 0.82), and total chest wall soft tissue (r = 0.85). CONCLUSION: Chest wall composition has a significant correlation with image noise for cardiac CT. Therefore, strategies that target radiation dose reduction should incorporate adaptation to chest wall composition. These determinations become more significant given the current obesity epidemic.

Quantification of arterial plaque and lumen density with MDCT.

PURPOSE: This study aimed to derive a mathematical correction function in order to normalize the CT number measurements for small volume arterial plaque and small vessel mimicking objects, imaged with multidetector CT (MDCT). METHODS: A commercially available calcium plaque phantom (QRM GmbH, Moehrendorf, Germany) and a custom built cardiovascular phantom were scanned with 320 and 64 MDCT scanners. The calcium hydroxyapatite plaque phantom contained objects 0.5-5.0 mm in diameter with known CT attenuation nominal values ranging 50-800 HU. The cardiovascular phantom contained vessel mimicking objects 1.0-5.0 mm in diameter with different contrast media. Both phantoms were scanned using clinical protocols for CT angiography and images were reconstructed with different filter kernels. The measured CT number (HU) and diameter of each object were analyzed on three clinical postprocessing workstations. From the resultant data, a mathematical formula was derived based on absorption function exp(--micro.-d) to demonstrate the relation between measured CT numbers and object diameters. RESULTS: The percentage reduction in measured CT number (HU) for the group of selected filter kernels, apparent during CT angiography, is dependent only on the object size (plaque or vessel diameter). The derived formula of the form 1-c.-exp(-a.-d--b) showed reduction in CT number for objects between 0.5 and 5 mm in diameter, with asymptote reaching background noise for small objects with diameters nearing the CT in-plane resolution (0.35 mm). No reduction was observed for the objects with diameters equal or larger than 5 mm. CONCLUSIONS: A clear mathematical relationship exists between object diameter and reduction in measured CT number in HU. This function is independent of exposure parameters and inherent attenuation properties of the objects studied. Future developments include the incorporation of this mathematical model function into quantification software in order to automatically generate a true assessment of measured CT number (HU) corresponding to plaque physical density rho (g/cm(3)). This is a significant development for the accurate, noninvasive classification of noncalcified arterial plaque.

Radiation dose optimisation in dynamic volume CT of the heart: tube current adaptation based on anterior-posterior chest diameter.

To compare tube current adaptation based on 3 body mass index (BMI) categories versus anterior-posterior chest diameter (APD) for radiation dose optimisation in patients undergoing dynamic volume cardiac CT. Two cardiac imaging centres participated in the study. 20 patients underwent a prospectively triggered 320-slice single beat cardiac CT using the X-ray tube current [mA] manually adjusted to the patient's BMI (group I). In 20 subsequent patients, the tube current was adapted according to the patient's APD (group II). All other parameters were kept constant. Image noise was defined as the standard deviation of attenuation values and measured using a ROI in the descending aorta. Variation in image noise was statistically compared between both patient groups. Average and standard deviation of pixel noise were 29.1 HU and 14.8 HU in group I and 28.0 HU and 4.2 HU in group II. Inter-individual variation of pixel noise was significantly lower in group II compared to group I (p < 0.0001). Tube current adaptation based on APD is superior to stepwise adaptation based on BMI for optimising radiation dose in dynamic volume cardiac CT and therefore limits unnecessary radiation dose while ensuring diagnostic image quality in patients with diverse body habitus.

Magnetic resonance imaging demonstration of a single lesion causing Wallerian degeneration in ascending and descending tracts in the spinal cord.

A magnetic resonance image of a 50-year-old man with a remote history of cervical spine injury showed focal myelomalacia at C5 and hyperintense areas on T2-weighted images laterally and posteriorly in the cord above and below C5. We believe these lesions to be due to Wallerian degeneration, with the cephalocaudal level of the Wallerian degeneration lesions dependant on the direction of the tracts relative to the C5 lesion.

Diagnostic performance of a prototype dual-energy chest imaging system ROC analysis.

RATIONALE AND OBJECTIVES: To assess the performance of an experimental prototype dual-energy (DE) chest imaging system in comparison to digital radiography (DR) in detection and characterization of lung lesions using receiver-operating characteristic (ROC) tests. MATERIALS AND METHODS: A cohort of 129 patients (80 M, 49 F; mean age, 64.8 years) was drawn from a trial of patients referred for percutaneous biopsy of a lung lesion. DR and DE images were acquired of each patient (posteroanterior view) before biopsy using a prototype system developed in our laboratory. The system incorporated a flat-panel detector and previously reported imaging techniques optimized such that the total dose for the DE image was equivalent to that of a DR acquisition. Each DE image was decomposed to three components (soft-tissue, bone, and composite "equivalent radiograph") by log subtraction with optimized noise reduction techniques. ROC tests were performed to evaluate the diagnostic performance of DR imaging in comparison to DE for nodule detection, with 258 left/right "half-chest" images derived from the 129 cases to give a roughly equal number of disease and normal cases. Five chest radiologists scored 258 half-chest DE and 258 half-chest DR (516 in total) images on a 5-point scale, and results (including ROC and area under the curve [AUC]) were analyzed using the ROCkit toolkit. Statistical significance in the observed differences was evaluated in terms of P values determined by a z test. Performance was analyzed for all cases pooled (258 DE vs. 258 DR images) and by retrospective stratification of the data according to nodule size, density, gender, lung region, and chest thickness. RESULTS: For results pooled over the entire cohort, there was no significant difference in ROC performance between DE and DR (AUC(DE) = 0.795 AUC(DR) = 0.789; P = .696). This finding is believed to be due to a large portion of lesions that were fairly conspicuous in either modality. In retrospective analysis of subgroups, a significant advantage was measured for DE imaging of small nodules (<1 cm diameter; AUC(DE) = 0.778; AUC(DR) = 0.706; P = .056), for nodules located in the right upper lobe (AUC(DE) = 0.836; AUC(DR) = 0.779; P = .003), and nodules located in right lower lobe (AUC(DE) = 0.804; AUC(DR) = 0.752; P = .054). DE imaging provided a clinically significant differential diagnosis in approximately one third of patients (49/158) (ie, disease cases in which the lesion was correctly identified in DE [(ROC rating > or =3], but missed in DR [ROC rating < or =2]). DE imaging also appeared to provide more definitive diagnosis (ie, a greater proportion of ROC ratings = 5 and 1 for identification of disease and normal cases, respectively), which presumably translates to increased confidence and a steeper ROC curve (even if the AUC are the same). CONCLUSIONS: DE imaging at dose equivalent to DR exhibited similar overall ROC performance to DR, although the radiologists noted qualitatively improved visualization (eg, improved characterization of lesion margins, visibility of calcifications and rib fractures). DE imaging demonstrated significant improvement in diagnostic performance for specific subgroups, including subcentimeter lung lesions and lesions in the right upper lobe, each of which is a potentially important factor in detecting early-stage malignancy.

Development of a high-performance dual-energy chest imaging system: initial investigation of diagnostic performance.

RATIONALE AND OBJECTIVES: The aim of this study was to assess the performance of a newly developed dual-energy (DE) chest radiographic system in comparison to digital radiographic (DR) imaging in the detection and characterization of lung nodules. MATERIALS AND METHODS: An experimental prototype was developed for high-performance DE chest imaging, with total dose equivalent to a single posterior-anterior DR image. Projections at low and high peak kilovoltage were used to decompose DE soft tissue and bone images. A cohort of 55 patients (31 men, 24 women; mean age, 65.6 years) was drawn from an ongoing trial involving patients referred for percutaneous computed tomography-guided biopsy of suspicious lung nodules. DE and DR images were acquired of each patient prior to biopsy. Image quality was assessed by means of human observer tests involving five radiologists independently rating the detection and characterization of lung nodules on a nine-point scale. Results were analyzed in terms of the fraction of cases at or above a given rating, and statistical significance was evaluated using Wilcoxon's signed-rank test. Performance was analyzed for all cases pooled as well as by stratification of nodule size, density, lung region, and chest thickness. RESULTS: The studies demonstrated a significant performance advantage for DE imaging compared to DR imaging (P < .001) in the detection and characterization of lung nodules. DE imaging improved the detection of both small and large nodules and exhibited the most significant improvement in regions of the upper lobes, where overlying anatomic noise (ribs and clavicles) are believed to reduce nodule conspicuity on DR imaging. CONCLUSIONS: DE imaging outperformed DR imaging overall, particularly in the detection of small, solid nodules. DE imaging also performed better in regions dominated by anatomic noise, such as the lung apices. The potential for improved nodule detection and characterization at radiation doses equivalent to DR imaging is encouraging and could augment the broader use of DE imaging. Future studies will extend the initial cohort and rating scale tests to a larger cohort evaluated by receiver-operating characteristic analysis and will evaluate DE imaging in comparison and as an adjuvant to low-dose computed tomography.

The use of computer-aided detection for the assessment of pulmonary arterial filling defects at computed tomographic angiography.

PURPOSE: To validate a computer-aided detection (CAD) tool for the detection of pulmonary arterial filling defects at computed tomographic pulmonary angiography (CTPA) and to assess its benefit for readers of different levels of experience. METHODS: One hundred consecutive CTPA studies were retrospectively evaluated by a chest radiologist for presence of emboli, serving as the reference standard. Subsequently, examinations were analyzed using commercially available second-generation CAD software (ImageChecker CT, version 2.1; R2 Technology, Inc., Sunnyvale, Calif). The staff radiologist assessed all CAD marks and classified them as true positive or false positive (FP), and any unmarked emboli were classified as false negative. Computer-aided detection software was also evaluated on a case basis compared with the reference standard.For the second part of the study, the 100 CTPAs were reviewed by 3 additional readers of different levels of experience, both without and with CAD, and findings correlated with the reference standard. RESULTS: Twenty-one studies (21%) were positive for pulmonary embolism. Of these, 18 were true positive on a case basis, and 3 were false negative. Of the 79 negative studies, 16 were true negative with no CAD marks, and the remaining 63 were FP. On a case basis, CAD sensitivity was 86%, specificity was 20%, negative predictive value was 84%, and positive predictive value (PPV) was 22%.Overall, the CAD software yielded 318 marks, identifying 64 of 93 emboli with an additional 254 FP marks. On a mark basis, sensitivity was 69%, and PPV was 20%.Computer-aided detection did not influence the most experienced reader (a chest fellow). Although CAD improved the subjective confidence of the second-year resident in some cases, it had no influence on overall interpretation or accuracy. Computer-aided detection improved accuracy only for the most inexperienced reader, helping this reader to identify 9 emboli not initially appreciated. CONCLUSIONS: Computer-aided detection specificity and PPV are poor due to expected FP marks, although, often, these can be easily dismissed. However, CAD software may play an important role as a second reader for residents or inexperienced readers.