Image-based automated Psoriasis Area Severity Index scoring by Convolutional Neural Networks.

BACKGROUND: The Psoriasis Area and Severity Index (PASI) score is commonly used in clinical practice and research to monitor disease severity and determine treatment efficacy. Automating the PASI score with deep learning algorithms, like Convolutional Neural Networks (CNNs), could enable objective and efficient PASI scoring. OBJECTIVES: To assess the performance of image-based automated PASI scoring in anatomical regions by CNNs and compare the performance of CNNs to image-based scoring by physicians. METHODS: Imaging series were matched to PASI subscores determined in real life by the treating physician. CNNs were trained using standardized imaging series of 576 trunk, 614 arm and 541 leg regions. CNNs were separately trained for each PASI subscore (erythema, desquamation, induration and area) in each anatomical region (trunk, arms and legs). The head region was excluded for anonymity. Additionally, PASI-trained physicians retrospectively determined image-based subscores on the test set images of the trunk. Agreement with the real-life scores was determined with the intraclass correlation coefficient (ICC) and compared between the CNNs and physicians. RESULTS: Intraclass correlation coefficients between the CNN and real-life scores of the trunk region were 0.616, 0.580, 0.580 and 0.793 for erythema, desquamation, induration and area, respectively, with similar results for the arms and legs region. PASI-trained physicians (N = 5) were in moderate-good agreement (ICCs 0.706-0.793) with each other for image-based PASI scoring of the trunk region. ICCs between the CNN and real-life scores were slightly higher for erythema (0.616 vs. 0.558), induration (0.580 vs. 0.573) and area scoring (0.793 vs. 0.694) than image-based scoring by physicians. Physicians slightly outperformed the CNN on desquamation scoring (0.580 vs. 0.589). CONCLUSIONS: Convolutional Neural Networks have the potential to automatically and objectively perform image-based PASI scoring at an anatomical region level. For erythema, desquamation and induration scoring, CNNs performed similar to physicians, while for area scoring CNNs outperformed physicians on image-based PASI scoring.

Medial closing wedge high tibial osteotomy for valgus tibial deformities: good clinical results and survival with a mean 4.5 years of follow-up in 113 patients.

PURPOSE: A varus-producing medial closing wedge high tibial osteotomy (MCWHTO) is an uncommon procedure. The aim of this retrospective study was to assess the survivorship and prevalence of post-operative subjective knee laxity and satisfaction in a large cohort of patients with a MCWHTO performed without a MCL-reefing procedure. METHODS: All patients (n = 176) who underwent a MCWHTO in our clinic between 2008 and 2016 were approached to participate. After review of patient charts, questionnaires were sent to willingly patients. Primary outcome was the survivorship of the MCWHTO; secondary outcome was patient-reported instability and satisfaction. RESULTS: One-hundred and thirteen patients participated in the study. The 5-year survival rate of the MCWHTO was almost 80%. A total of 77% of the patients was satisfied with the treatment. With regard to post-operative subjective knee laxity, 26% of the patients experienced instability of the knee post-operation. Instability was significantly correlated with the KOOS domains, the Lysholm score, the IKDC knee function score and the Physical and Mental Health Domains of the SF-36. CONCLUSION: Medial closing wedge high tibial osteotomy provides good results regarding survivorship and patient satisfaction for patients with a valgus deformity which is located in the proximal tibia. Clinically relevant is that in the surgical technique without MCL-reefplasty instability is significantly correlated with worse patient-reported outcome measures. The addition of a MCL reefing procedure will improve outcome in selected patients. LEVEL OF EVIDENCE: III.

Chest radiography: new technological developments and their applications.

Digital chest radiography is still the most common radiological examination. With the upcoming three-dimensional (3D) acquisition techniques the value of radiography seems to diminish. But because radiography is inexpensive, readily available, and requires very little dose, it is still being used for the first-line detection of many cardiothoracic diseases. In the last decades major technical developments of this 2D technique are being achieved. First, hardware developments of digital radiography have improved the contrast to noise, dose efficacy, throughput, and workflow. Dual energy acquisition techniques reduce anatomical noise by splitting a chest radiograph into a soft tissue image and a bone image. Second, advanced processing methods are developed to enable and improve detection of many kinds of disease. Digital bone subtraction by a software algorithm mimics the soft tissue image normally acquired with dedicated hardware. Temporal subtraction aims to rule out anatomical structures clotting the image, by subtracting a current radiograph with a previous radiograph. Finally, computer-aided detection systems help radiologists for the detection of various kinds of disease such as pulmonary nodules or tuberculosis.

Detecting Mandible Fractures in CBCT Scans Using a 3-Stage Neural Network.

After nasal bone fractures, fractures of the mandible are the most frequently encountered injuries of the facial skeleton. Accurate identification of fracture locations is critical for effectively managing these injuries. To address this need, JawFracNet, an innovative artificial intelligence method, has been developed to enable automated detection of mandibular fractures in cone-beam computed tomography (CBCT) scans. JawFracNet employs a 3-stage neural network model that processes 3-dimensional patches from a CBCT scan. Stage 1 predicts a segmentation mask of the mandible in a patch, which is subsequently used in stage 2 to predict a segmentation of the fractures and in stage 3 to classify whether the patch contains any fracture. The final output of JawFracNet is the fracture segmentation of the entire scan, obtained by aggregating and unifying voxel-level and patch-level predictions. A total of 164 CBCT scans without mandibular fractures and 171 CBCT scans with mandibular fractures were included in this study. Evaluation of JawFracNet demonstrated a precision of 0.978 and a sensitivity of 0.956 in detecting mandibular fractures. The current study proposes the first benchmark for mandibular fracture detection in CBCT scans. Straightforward replication is promoted by publicly sharing the code and providing access to JawFracNet on grand-challenge.org.

AI-support for the detection of intracranial large vessel occlusions: One-year prospective evaluation.

Purpose: Few studies have evaluated real-world performance of radiological AI-tools in clinical practice. Over one-year, we prospectively evaluated the use of AI software to support the detection of intracranial large vessel occlusions (LVO) on CT angiography (CTA). Method: Quantitative measures (user log-in attempts, AI standalone performance) and qualitative data (user surveys) were reviewed by a key-user group at three timepoints. A total of 491 CTA studies of 460 patients were included for analysis. Results: The overall accuracy of the AI-tool for LVO detection and localization was 87.6%, sensitivity 69.1% and specificity 91.2%. Out of 81 LVOs, 31 of 34 (91%) M1 occlusions were detected correctly, 19 of 38 (50%) M2 occlusions, and 6 of 9 (67%) ICA occlusions. The product was considered user-friendly. The diagnostic confidence of the users for LVO detection remained the same over the year. The last measured net promotor score was -56%. The use of the AI-tool fluctuated over the year with a declining trend. Conclusions: Our pragmatic approach of evaluating the AI-tool used in clinical practice, helped us to monitor the usage, to estimate the perceived added value by the users of the AI-tool, and to make an informed decision about the continuation of the use of the AI-tool.

The relationship between lung function impairment and quantitative computed tomography in chronic obstructive pulmonary disease.

OBJECTIVES: To determine the relationship between lung function impairment and quantitative computed tomography (CT) measurements of air trapping and emphysema in a population of current and former heavy smokers with and without airflow limitation. METHODS: In 248 subjects (50 normal smokers; 50 mild obstruction; 50 moderate obstruction; 50 severe obstruction; 48 very severe obstruction) CT emphysema and CT air trapping were quantified on paired inspiratory and end-expiratory CT examinations using several available quantification methods. CT measurements were related to lung function (FEV(1), FEV(1)/FVC, RV/TLC, Kco) by univariate and multivariate linear regression analysis. RESULTS: Quantitative CT measurements of emphysema and air trapping were strongly correlated to airflow limitation (univariate r-squared up to 0.72, p < 0.001). In multivariate analysis, the combination of CT emphysema and CT air trapping explained 68-83% of the variability in airflow limitation in subjects covering the total range of airflow limitation (p < 0.001). CONCLUSIONS: The combination of quantitative CT air trapping and emphysema measurements is strongly associated with lung function impairment in current and former heavy smokers with a wide range of airflow limitation.

Quantitative computed tomography in COPD: possibilities and limitations.

Chronic obstructive pulmonary disease (COPD) is a heterogeneous disease that is characterized by chronic airflow limitation. Unraveling of this heterogeneity is challenging but important, because it might enable more accurate diagnosis and treatment. Because spirometry cannot distinguish between the different contributing pathways of airflow limitation, and visual scoring is time-consuming and prone to observer variability, other techniques are sought to start this phenotyping process. Quantitative computed tomography (CT) is a promising technique, because current CT technology is able to quantify emphysema, air trapping, and large airway wall dimensions. This review focuses on CT quantification techniques of COPD disease components and their current status and role in phenotyping COPD.

Automated detection of cerebral microbleeds via segmentation in susceptibility-weighted images of patients with traumatic brain injury.

Cerebral microbleeds (CMBs) are a recognised biomarker of traumatic axonal injury (TAI). Their number and location provide valuable information in the long-term prognosis of patients who sustained a traumatic brain injury (TBI). Accurate detection of CMBs is necessary for both research and clinical applications. CMBs appear as small hypointense lesions on susceptibility-weighted magnetic resonance imaging (SWI). Their size and shape vary markedly in cases of TBI. Manual annotation of CMBs is a difficult, error-prone, and time-consuming task. Several studies addressed the detection of CMBs in other neuropathologies with convolutional neural networks (CNNs). In this study, we developed and contrasted a classification (Patch-CNN) and two segmentation (Segmentation-CNN, U-Net) approaches for the detection of CMBs in TBI cases. The models were trained using 45 datasets, and the best models were chosen according to 16 validation sets. Finally, the models were evaluated on 10 TBI and healthy control cases, respectively. Our three models outperform the current status quo in the detection of traumatic CMBs, achieving higher sensitivity at low false positive (FP) counts. Furthermore, using a segmentation approach allows for better precision. The best model, the U-Net, achieves a detection rate of 90% at FP counts of 17.1 in TBI patients and 3.4 in healthy controls.

Association of the transfer coefficient of the lung for carbon monoxide with emphysema progression in male smokers.

A decreased transfer coefficient of the lung for carbon monoxide (K(CO)) is associated with emphysema. We evaluated whether in heavy smokers, baseline K(CO) was associated with the progression of computed tomography (CT)-detected emphysema, and the progression of airflow limitation. Heavy smokers, mean ± sd 41.3 ± 18.7 pack-yrs, participating in a lung cancer screening trial underwent diffusion testing and CT scanning of the lungs. CT scanning was repeated after median (25th-75th percentile) 2.8 (2.7-3.0) yrs and emphysema was assessed by lung densitometry using the 15th percentile. The association between K(CO) at baseline with progression of emphysema and lung function decline was assessed by multiple linear regression, correcting for baseline CT-quantified emphysema severity and forced expiratory volume in 1 s (FEV1/forced vital capacity (FVC), age, height, body mass index, pack-yrs and smoking status (current or former smoker). 522 participants aged 60.1 ± 5.4 yrs were included. Mean ± sd 15th percentile was -938 ± 19, absolute FEV1/FVC was 71.6 ± 9% and K(CO) was 1.23 ± 0.25, which is 81.8 ± 16.5% of predicted. By interpolation, a one sd (0.25) lower K(CO) value at baseline predicted a 1.6 HU lower 15th percentile and a 0.78% lower FEV1/FVC after follow-up (p < 0.001). A lower baseline K(CO) value is independently associated with a more rapid progression of emphysema and airflow limitation in heavy smokers.

Modified Chrispin-Norman chest radiography score for cystic fibrosis: observer agreement and correlation with lung function.

OBJECTIVE: To test observer agreement and two strategies for possible improvement (consensus meeting and reference images) for the modified Chrispin-Norman score for children with cystic fibrosis (CF). METHODS: Before and after a consensus meeting and after developing reference images three observers scored sets of 25 chest radiographs from children with CF. Observer agreement was tested for line, ring, mottled and large soft shadows, for overinflation and for the composite modified Chrispin-Norman score. Correlation with lung function was assessed. RESULTS: Before the consensus meeting agreement between observers 1 and 2 was moderate-good, but with observer 3 agreement was poor-fair. Scores correlated significantly with spirometry for observers 1 and 2 (-0.72

Automated chest X-ray reading for tuberculosis in the Philippines to improve case detection: a cohort study.

BACKGROUND: DetecTB (Diagnostic Enhanced Tools for Extra Cases of TB), an intensified tuberculosis (TB) case-finding programme targeting prisons and high-risk communities was implemented on Palawan Island, the Philippines.OBJECTIVE: To evaluate the performance of TB detection based on computerised chest radiography (CXR) readings.DESIGN: Data from 14 094 subjects were analysed from September 2012 to June 2014. All CXRs were read by a physician and by software. Individuals with TB symptoms or CXR abnormalities according to the physician underwent Xpert® MTB/RIF testing, the remaining persons were considered TB-negative (screening reference). A subset of 200 CXRs was read by an independent human reader (radiological reference). This reader also re-read a subset of the most abnormal cases as identified using the software but read as normal by the physician (discordant cases).RESULTS: A total of 10 755 individuals were included in the analysis, 2534 of whom had a positively assessed CXR; 298 cases were Xpert-positive. Using the screening reference, the area under the receiver operating characteristic curve for software readings was 0.93 (95%CI 0.92-0.94), with a sensitivity of 0.98 (95%CI 0.97-0.99) and a specificity of 0.69 (95%CI 0.40-0.98). Based on the radiological reference, the physician performed slightly worse than the software (sensitivity, 0.82, 95%CI 0.74-0.89 and specificity, 0.87, 95%CI 0.81-0.96 vs. sensitivity, 0.83, 95%CI 0.71-0.93 and specificity, 0.87, 95%CI 0.75-0.95), although this was not statistically significant. Of the 291 discordant cases, 70% were assessed as positive, resulting in a 22% increase in TB detection when extrapolated to the full cohort.CONCLUSION: The performance of automated CXR reading is comparable to that of the attending physicians in DetecTB, and its use as a second reader could increase TB detection.

Supervised enhancement filters: application to fissure detection in chest CT scans.

In medical image processing, many filters have been developed to enhance certain structures in 3-D data. In this paper, we propose to use pattern recognition techniques to design more optimal filters. The essential difference with previous approaches is that we provide a system with examples of what it should enhance and suppress. This training data is used to construct a classifier that determines the probability that a voxel in an unseen image belongs to the target structure(s). The output of a rich set of basis filters serves as input to the classifier. In a feature selection process, this set is reduced to a compact, efficient subset. We show that the output of the system can be reused to extract new features, using the same filters, that can be processed by a new classifier. Such a multistage approach further improves performance. While the approach is generally applicable, in this work the focus is on enhancing pulmonary fissures in 3-D computed tomography (CT) chest scans. A supervised fissure enhancement filter is evaluated on two data sets, one of scans with a normal clinical dose and one of ultra-low dose scans. Results are compared with those of a recently proposed conventional fissure enhancement filter. It is demonstrated that both methods are able to enhance fissures, but the supervised approach shows better performance; the areas under the receiver operating characteristic (ROC) curve are 0.98 versus 0.90, for the normal dose data and 0.97 versus 0.87 for the ultra low dose data, respectively.

Medial collateral ligament laxity in valgus knee deformity before and after medial closing wedge high tibial osteotomy measured with instrumented laxity measurements and patient reported outcome.

INTRODUCTION: Medial closing wedge high tibial osteotomy (CWHTO) for valgus deformity correction was first described by Coventry whom performed an additional reefing of the medial collateral ligament (MCL) to prevent instability postoperative. In our clinic the additional reefing procedure has never been performed and instability has not been reported routinely by patients. Using instrumented laxity testing, pre- and postoperative valgus and varus knee laxity can be measured objectively. We hypothesize that absence of changes in laxity testing and subjective knee stability scores support that no additional reefing procedure is necessary. MATERIALS AND METHODS: In a prospective cohort study 11 consecutive patients indicated for medial CWHTO were subjected to pre- and postoperative stress X-rays in 30° and 70° of flexion and opening of the joint line was measured in degrees on the radiographs. Patient reported outcome scores were documented with the KOOS, Lysholm, SF36, Oxford Knee Score and a VAS instability scoring tool. RESULTS: All patients (7 females) completed the study, mean age was 46 years. Mean preoperative Hip Knee Ankle angle 6.4° valgus was corrected to mean postoperative alignment 0.1° valgus. A significant difference was measured between mean pre- and postoperative 30° valgus laxity (2.8° vs 5.3°, P = 0.005), 30° varus laxity (6.7° vs 3.2°, P = 0.005) and 70° valgus laxity (2.0° vs 4.8°, P = 0.008). Postoperative patient-reported knee instability as measured with the Lysholm questionnaire was significantly improved compared to preoperative instability (P = 0.006). VAS instability improved, but didn't reach significance (8.0 preoperative and 5.5 postoperative (P = 0.127). Other outcome measures showed improvement as well. No correlations between radiological findings and outcome scores were found. CONCLUSION: A significant increase in postoperative valgus laxity in 30° and 70° of flexion deems reconsidering addition of MCL reefingplasty to the medial CWHTO although patient reported outcome on subjective stability scores fails to report increase of instability in this study population. Instrumented laxity measurements of medial CWHTO patients treated with additional medial reefingplasty should be performed to prove the value of this procedure.

Small airway segmentation in thoracic computed tomography scans: a machine learning approach.

Small airway obstruction is a main cause for chronic obstructive pulmonary disease (COPD). We propose a novel method based on machine learning to extract the airway system from a thoracic computed tomography (CT) scan. The emphasis of the proposed method is on including the smallest airways that are still visible on CT. We used an optimized sampling procedure to extract airway and non-airway voxel samples from a large set of scans for which a semi-automatically constructed reference standard was available. We created a set of features which represent tubular and texture properties that are characteristic for small airway voxels. A random forest classifier was used to determine for each voxel if it belongs to the airway class. Our method was validated on a set of 20 clinical thoracic CT scans from the COPDGene study. Experiments show that our method is effective in extracting the full airway system and in detecting a large number of small airways that were missed by the semi-automatically constructed reference standard.

Accuracy of an automated system for tuberculosis detection on chest radiographs in high-risk screening.

SETTING: Tuberculosis (TB) screening programmes can be optimised by reducing the number of chest radiographs (CXRs) requiring interpretation by human experts. OBJECTIVE: To evaluate the performance of computerised detection software in triaging CXRs in a high-throughput digital mobile TB screening programme. DESIGN: A retrospective evaluation of the software was performed on a database of 38 961 postero-anterior CXRs from unique individuals seen between 2005 and 2010, 87 of whom were diagnosed with TB. The software generated a TB likelihood score for each CXR. This score was compared with a reference standard for notified active pulmonary TB using receiver operating characteristic (ROC) curve and localisation ROC (LROC) curve analyses. RESULTS: On ROC curve analysis, software specificity was 55.71% (95%CI 55.21-56.20) and negative predictive value was 99.98% (95%CI 99.95-99.99), at a sensitivity of 95%. The area under the ROC curve was 0.90 (95%CI 0.86-0.93). Results of the LROC curve analysis were similar. CONCLUSION: The software could identify more than half of the normal images in a TB screening setting while maintaining high sensitivity, and may therefore be used for triage.

Computer-assisted chest radiography reading for tuberculosis screening in people living with diabetes mellitus.

BACKGROUND: Diabetes mellitus is a significant risk factor for tuberculosis (TB). We evaluated the performance of computer-aided detection for tuberculosis (CAD4TB) in people living with diabetes mellitus (PLWD) in Indonesia. METHODS: PLWD underwent symptom screening and chest X-ray (CXR); sputum was examined in those with positive symptoms and/or CXR. Digital CXRs were scored using CAD4TB and analysed retrospectively using clinical and microbiological diagnosis as a reference. The area under the receiver operator curve (AUC) of CAD4TB scores was determined, and an optimal threshold score established. Agreement between CAD4TB and the radiologist's reading was determined. RESULTS: Among 346 included PLWD, seven (2.0%) had microbiologically confirmed and two (0.6%) had clinically diagnosed TB. The highest agreement of CAD4TB with radiologist reading was achieved using a threshold score of 70 (κ = 0.41, P < 0.001). The AUC for CAD4TB was 0.89 (95%CI 0.73-1.00). A threshold score of 65 for CAD4TB resulted in a sensitivity, specificity, positive predictive value and negative predictive value of respectively 88.9% (95%CI 51.8-99.7), 88.5% (95%CI 84.6-91.7), 17.0% (95%CI 7.6-30.8) and 99.6% (95%CI 98.2-100). With this threshold, 48 (13.9%) individuals needed microbiological examination and no microbiologically confirmed cases were missed. CONCLUSIONS: CAD4TB has potential as a triage tool for TB screening in PLWD, thereby significantly reducing the need for microbiological examination.

Automatic versus human reading of chest X-rays in the Zambia National Tuberculosis Prevalence Survey.

SETTING: Tuberculosis (TB) prevalence survey in Zambia between 2013 and 2014. OBJECTIVE: To compare the performance of automatic software (CAD4TB 5) in chest X-ray (CXR) reading with that of field (general practitioners) and central (radiologists) readers. DESIGN: A retrospective study comparing the performance of human and automatic reading was conducted. Two scenarios for central reading were evaluated: abnormalities not consistent with TB were considered to be 'normal' or 'abnormal'. Sputum culture was defined as the reference standard. Measures derived from receiver operating characteristic analysis were used to assess readers' performances. RESULTS: Of 46 099 participants, 23 838 cases included all survey information; of these, 106 cases were culture-confirmed TB-positive. The performance of CAD4TB 5 was similar to that of field and central readers. Although there were significant differences in specificity when compared with field readings (P = 0.002) and central readings considering the first scenario (P < 0.001), these differences were not substantial (93.2% vs. 92.6% and 98.4% vs. 99.6%, respectively).CONCLUSIONp: The performance of automatic CXR readings is comparable with that of human experts in a TB prevalence survey setting using culture as reference.

Fast interactive segmentation of the pulmonary lobes from thoracic computed tomography data.

Automated lung lobe segmentation methods often fail for challenging and clinically relevant cases with incomplete fissures or substantial amounts of pathology. We present a fast and intuitive method to interactively correct a given lung lobe segmentation or to quickly create a lobe segmentation from scratch based on a lung mask. A given lobar boundary is converted into a mesh by principal component analysis of 3D lobar boundary markers to obtain a plane where nodes correspond to the position of the markers. An observer can modify the mesh by drawing on 2D slices in arbitrary orientations. After each drawing, the mesh is immediately adapted in a 3D region around the user interaction. For evaluation we participated in the international lung lobe segmentation challenge LObe and lung analysis 2011 (LOLA11). Two observers applied the method to correct a given lung lobe segmentation obtained by a fully automatic method for all 55 CT scans of LOLA11. On average observer 1/2 required 8 ± 4/25 ± 12 interactions per case and took 1:30 ± 0:34/3:19 ± 1:29 min. The average distances to the reference segmentation were improved from an initial 2.68 ± 14.71 mm to 0.89 ± 1.63/0.74 ± 1.51 mm. In addition, one observer applied the proposed method to create a segmentation from scratch. This took 3:44 ± 0:58 minutes on average per case, applying an average of 20 ± 3 interactions to reach an average distance to the reference of 0.77 ± 1.14 mm. Thus, both the interactive corrections and the creation of a segmentation from scratch were feasible in a short time with excellent results and only little interaction. Since the mesh adaptation is independent of image features, the method can successfully handle patients with severe pathologies, provided that the human operator is capable of correctly indicating the lobar boundaries.

Corrigendum: Towards automatic pulmonary nodule management in lung cancer screening with deep learning.

This corrects the article DOI: 10.1038/srep46479.

Filter learning: application to suppression of bony structures from chest radiographs.

A novel framework for image filtering based on regression is presented. Regression is a supervised technique from pattern recognition theory in which a mapping from a number of input variables (features) to a continuous output variable is learned from a set of examples from which both input and output are known. We apply regression on a pixel level. A new, substantially different, image is estimated from an input image by computing a number of filtered input images (feature images) and mapping these to the desired output for every pixel in the image. The essential difference between conventional image filters and the proposed regression filter is that the latter filter is learned from training data. The total scheme consists of preprocessing, feature computation, feature extraction by a novel dimensionality reduction scheme designed specifically for regression, regression by k-nearest neighbor averaging, and (optionally) iterative application of the algorithm. The framework is applied to estimate the bone and soft-tissue components from standard frontal chest radiographs. As training material, radiographs with known soft-tissue and bone components, obtained by dual energy imaging, are used. The results show that good correlation with the true soft-tissue images can be obtained and that the scheme can be applied to images from a different source with good results. We show that bone structures are effectively enhanced and suppressed and that in most soft-tissue images local contrast of ribs decreases more than contrast between pulmonary nodules and their surrounding, making them relatively more pronounced.