Crowd-Sourced Deep Learning for Intracranial Hemorrhage Identification: Wisdom of Crowds or Laissez-Faire

BACKGROUND AND PURPOSE

Researchers and clinical radiology practices are increasingly faced with the task of selecting the most accurate artificial intelligence tools from an ever-expanding range. In this study, we sought to test the utility of ensemble learning for determining the best combination from 70 models trained to identify intracranial hemorrhage. Furthermore, we investigated whether ensemble deployment is preferred to use of the single best model. It was hypothesized that any individual model in the ensemble would be outperformed by the ensemble.

MATERIALS AND METHODS

In this retrospective study, de-identified clinical head CT scans from 134 patients were included. Every section was annotated with "no-intracranial hemorrhage" or "intracranial hemorrhage," and 70 convolutional neural networks were used for their identification. Four ensemble learning methods were researched, and their accuracies as well as receiver operating characteristic curves and the corresponding areas under the curve were compared with those of individual convolutional neural networks. The areas under the curve were compared for a statistical difference using a generalized U-statistic.

RESULTS

The individual convolutional neural networks had an average test accuracy of 67.8% (range, 59.4%-76.0%). Three ensemble learning methods outperformed this average test accuracy, but only one achieved an accuracy above the 95th percentile of the individual convolutional neural network accuracy distribution. Only 1 ensemble learning method achieved a similar area under the curve as the single best convolutional neural network (Δarea under the curve = 0.03; 95% CI, -0.01-0.06; P = .17).

CONCLUSIONS

None of the ensemble learning methods outperformed the accuracy of the single best convolutional neural network, at least in the context of intracranial hemorrhage detection.

Personalized biomechanical tongue models based on diffusion-weighted MRI and validated using optical tracking of range of motion

For advanced tongue cancer, the choice between surgery and organ-sparing treatment is often dependent on the expected loss of tongue functionality after treatment. Biomechanical models might assist in this choice by simulating the post-treatment function loss. However, this function loss varies between patients and should, therefore, be predicted for each patient individually. In the present study, the goal was to better predict the postoperative range of motion (ROM) of the tongue by personalizing biomechanical models using diffusion-weighted MRI and constrained spherical deconvolution reconstructions of tongue muscle architecture. Diffusion-weighted MRI scans of ten healthy volunteers were obtained to reconstruct their tongue musculature, which were subsequently registered to a previously described population average or atlas. Using the displacement fields obtained from the registration, the segmented muscle fiber tracks from the atlas were morphed back to create personalized muscle fiber tracks. Finite element models were created from the fiber tracks of the atlas and those of the individual tongues. Via inverse simulation of a protruding, downward, left and right movement, the ROM of the tongue was predicted. This prediction was compared to the ROM measured with a 3D camera. It was demonstrated that biomechanical models with personalized muscles bundles are better in approaching the measured ROM than a generic model. However, to achieve this result a correction factor was needed to compensate for the small magnitude of motion of the model. Future versions of these models may have the potential to improve the estimation of function loss after treatment for advanced tongue cancer.

In-vivo tongue stiffness measured by aspiration: Resting vs general anesthesia

Tongue cancer treatment often results in impaired speech, swallowing, or mastication. Simulating the effect of treatments can help the patient and the treating physician to understand the effects and impact of the intervention. To simulate deformations of the tongue, identifying accurate mechanical properties of tissue is essential. However, not many succeeded in characterizing in-vivo tongue stiffness. Those who did, measured the tongue At Rest (AR), in which muscle tone subsides even if muscles are not willingly activated. We expected to find an absolute rest state in participants 'under General Anesthesia' (GA). We elaborated on previous work by measuring the mechanical behavior of the in-vivo tongue under aspiration using an improved volume-based method. Using this technique, 5 to 7 measurements were performed on 10 participants both AR and under GA. The obtained Pressure-Shape curves were first analyzed using the initial slope and its variations. Hereafter, an inverse Finite Element Analysis (FEA) was applied to identify the mechanical parameters using the Yeoh, Gent, and Ogden hyperelastic models. The measurements AR provided a mean Young's Modulus of 1638 Pa (min 1035 - max 2019) using the Yeoh constitutive model, which is in line with previous ex-vivo measurements. However, while hoping to find a rest state under GA, the tongue unexpectedly appeared to be approximately 2 to 2.5 times stiffer under GA than AR. Explanations for this were sought by examining drugs administered during GA, blood flow, perfusion, and upper airway reflexes, but neither of these explanations could be confirmed.

Interactive 3D U-net for the segmentation of the pancreas in computed tomography scans

The increasing incidence of pancreatic cancer will make it the second deadliest cancer in 2030. Imaging based early diagnosis and image guided treatment are emerging potential solutions. Artificial intelligence (AI) can help provide and improve widespread diagnostic expertise and accurate interventional image interpretation. Accurate segmentation of the pancreas is essential to create annotated data sets to train AI, and for computer assisted interventional guidance. Automated deep learning segmentation performance in pancreas computed tomography (CT) imaging is low due to poor grey value contrast and complex anatomy. A good solution seemed a recent interactive deep learning segmentation framework for brain CT that helped strongly improve initial automated segmentation with minimal user input. This method yielded no satisfactory results for pancreas CT, possibly due to a sub-optimal neural network architecture. We hypothesize that a state-of-the-art U-net neural network architecture is better because it can produce a better initial segmentation and is likely to be extended to work in a similar interactive approach. We implemented the existing interactive method, iFCN, and developed an interactive version of U-net method we call iUnet. The iUnet is fully trained to produce the best possible initial segmentation. In interactive mode it is additionally trained on a partial set of layers on user generated scribbles. We compare initial segmentation performance of iFCN and iUnet on a 100CT dataset using dice similarity coefficient analysis. Secondly, we assessed the performance gain in interactive use with three observers on segmentation quality and time. Average automated baseline performance was 78% (iUnet) versus 72% (FCN). Manual and semi-automatic segmentation performance was: 87% in 15 min. for manual, and 86% in 8 min. for iUNet. We conclude that iUnet provides a better baseline than iFCN and can reach expert manual performance significantly faster than manual segmentation in case of pancreas CT. Our novel iUnet architecture is modality and organ agnostic and can be a potential novel solution for semi-automatic medical imaging segmentation in general.

Quantification of tongue mobility impairment using optical tracking in patients after receiving primary surgery or chemoradiation

PURPOSE

Tongue mobility has shown to be a clinically interesting parameter on functional results after tongue cancer treatment which can be objectified by measuring the Range Of Motion (ROM). Reliable measurements of ROM would enable us to quantify the severity of functional impairments and use these for shared decision making in treatment choices, rehabilitation of speech and swallowing disturbances after treatment.

METHOD

Nineteen healthy participants, eighteen post-chemotherapy patients and seventeen post-surgery patients were asked to perform standardized tongue maneuvers in front of a 3D camera system, which were subsequently tracked and corrected for head and jaw motion. Indicators, such as the left-right tongue range and the deflection angle with the horizontal axis were extracted from the tongue trajectory to serve as a quantitative measure for the impaired tongue mobility.

RESULTS

The range and deflection angle showed an excellent intra- and interrater reliability (ICC 0.9) The repeatability experiment showed an average standard deviation of 2.5 mm to 3.5 mm for every movement, except the upward movement. The post-surgery patient group showed a smaller tongue range and higher deflection angle overall than the healthy participants. Post-chemoradiation patients showed less difference in tongue ROM compared with healthy participants. Only a few patients showed asymmetrical movement after treatment, which could not always be explained by T-stage or the side of treatment alone.

CONCLUSION

We introduced a reliable and reproducible method for measuring the ROM and to quantify for motion impairments, that was able to show differences in tongue ROM between healthy subjects and patients after chemoradiation or surgery. Future research should focus on measuring patients with oral cancer pre- and post-treatment in combination with the collection of detailed information about the individual tongue anatomy, so that the full ROM trajectory can be used to identify changes over time and to quantify functional impairment.

An interactive surgical simulation tool to assess the consequences of a partial glossectomy on a biomechanical model of the tongue

Oral cancer surgery has a negative influence on the quality of life (QOL). As a result of the complex physiology involved in oral functions, estimation of surgical effects on functionality remains difficult. We present a user-friendly biomechanical simulation of tongue surgery, including closure with suturing and scar formation, followed by an automated adaptation of a finite element (FE) model to the shape of the tongue. Different configurations of our FE model were evaluated and compared to a well-established FE model. We showed that the post-operative impairment as predicted by our model was qualitatively comparable to a patient case for five different tongue maneuvers.

Validation of low-cost smartphone-based thermal camera for diabetic foot assessment

AIMS

Infrared thermal imaging (IR) is not yet routinely implemented for early detection of diabetic foot ulcers (DFU), despite proven clinical effectiveness. Low-cost, smartphone-based IR-cameras are now available and may lower the threshold for implementation, but the quality of these cameras is unknown. We aim to validate a smartphone-based IR-camera against a high-end IR-camera for diabetic foot assessment.

METHODS

We acquired plantar IR images of feet of 32 participants with a current or recently healed DFU with the smartphone-based FLIR-One and the high-end FLIR-SC305. Contralateral temperature differences of the entire plantar foot and nine pre-specified regions were compared for validation. Intra-class correlations coefficient (ICC(3,1)) and Bland-Altman plots were used to test agreement. Clinical validity was assessed by calculating statistical measures of diagnostic performance.

RESULTS

Almost perfect agreement was found for temperature measurements in both the entire plantar foot and the combined pre-specified regions, respectively, with ICC values of 0.987 and 0.981, Bland-Altman plots' mean Δ = -0.14 and Δ = -0.06. Diagnostic accuracy showed 94% and 93% sensitivity, and 86% and 91% specificity.

CONCLUSIONS

The smartphone-based IR-camera shows excellent validity for diabetic foot assessment.

Feasibility of automated target centralization in colonoscopy

Purpose

Early detection of colorectal cancer is key to full recovery. This urged governments to start population screening programs for colorectal cancer, often using flexible endoscopes. Flexible endoscopy is difficult to learn and time-consuming. Automation of flexible endoscopes may increase the capacity for the screening programs. The goal of this pilot study is to investigate the clinical and technical feasibility of an assisting automated navigation algorithm for a colonoscopy procedure.

Methods

Automated navigation (lumen centralization) was implemented in a robotized system designed for conventional flexible endoscopes. Ten novice and eight expert users were asked to perform a diagnostic colonoscopy on a colon model twice: once using the conventional and once using the robotic system. Feasibility was evaluated using time and location data as measures of the system’s added value.

Results

Automated target centralization (ATC) was turned on by the novices for a median of 4.2 % of the time during insertion and 0.3 % during retraction. Experts turned ATC on for 4.0 % of the time during insertion and 11.6 % during retraction. Novices and experts showed comparable times to reach the cecum with the conventional or the robotic setup with ATC.

Conclusion

The ATC algorithm combined with the robotized endoscope setup works in an experimental setup that closely resembles the clinical environment and is considered feasible, although ATC use was lower than expected. For novices, it was unclear whether the low usage was due to unfamiliarity with the system or because they did not need ATC. Experts used ATC also during the retraction phase of the procedure. This was an unexpected finding and may indicate an added value of the system.

Electronic supplementary material

The online version of this article (doi:10.1007/s11548-015-1301-3) contains supplementary material, which is available to authorized users.

Make it spin: individual trapping of sperm for analysis and recovery using micro-contact printing

In this article, we describe the development of a high throughput platform to spatially manipulate viable sperm for motility measurements and recovery of the best single sperm for fertilization purposes. Micro-contact printing was used to pattern islands of adhesive proteins (fibronectin) separated by sperm repellent species (Pluronic acid F-127) on commercially available polystyrene substrates. Following washing, arrays of viable single sperm were captured onto the islands demonstrating for the first time that sperm can be trapped by micro-contact printing with patterning efficiency of 90% while retaining 100% viability. These were then subjected to motility analysis whilst remaining spatially confined to the islands. Single sperm motility was assessed (n = 37) by software analysis measuring the number of rotations per second (degrees s⁻¹). The assignment of array coordinates allows the more active single sperm to be easily identified and recovered by a simple micromanipulator pipette aspiration step with automated possibility for assisted reproductive technologies or further quality correlation analysis. Taken together, we show for the first time a technique to simultaneously screen thousands of viable single sperm for motility assessment while retaining the ability for single species recovery for enhanced fertilization purposes.

Low rates of treatment of cardiovascular risk factors in patients treated with antipsychotics

Introduction

Cardiovascular disease is the most common cause of death in patients with severe mental illness (SMI). Patients are more likely to be overweight and to have hypertension, dyslipidemia, hyperglycemia and diabetes mellitus. Antipsychotic medication can induce or worsen these cardiovascular risk factors.

Objectives

Accessibility to healthcare and adequate somatic treatment is limited in patients with SMI.

Aims

Prevalence of inadequate treatment for metabolic cardiovascular risk factors in patients treated with antipsychotic medication.

Methods

A health monitor was introduced as a screening instrument in a schizophrenia treatment and recovery program (so called F-ACT) at the outpatient departments. Assesment included physical and laboratory examination, demographics, DSM-IV diagnoses, remission-criteria, social functioning, use of medication and drugs.

Results

Over a period of 18 months (2009–2010), 520 of the 600 patients included in the F-ACT were evaluated. Preliminary analysis shows that 452 patients met DSM-IV criteria of schizophrenia and other psychotic disorders. All patients were treated with at least one antipsychotic agent. Metabolic syndrome according to ATP-III was present in 50% of patients (60% abdominal obesity, 58% hypertension, 24% hyperglycemia, 52% high triglycerides, 49% low HDL-cholesterol). Rates of non-treatment ranged from 78% for hypertension, 85% for dyslipidemia and 48% for diabetes mellitus. Rates of inadequate treatment ranged from 69% for hypertension and 83% for dyslipidemia.

Conclusions

1.

Consistency checks for particle filters

An "inconsistent" particle filter produces--in a statistical sense--larger estimation errors than predicted by the model on which the filter is based. Two test variables are introduced that allow the detection of inconsistent behavior. The statistical properties of the variables are analyzed. Experiments confirm their suitability for inconsistency detection.

An immunocytochemical double staining technique applied to intraepithelial lymphocytes in the small intestine of the rat

To study lymphocyte antigens in cryostat sections of small intestine of the rat, an indirect double labelling technique was developed, avoiding alkaline phosphatase for the 2nd label, in view of the high endogenous activity of this enzyme in the enterocytes. 2 times a peroxidase was used, but in the 1st sequence an antibody conjugated peroxidase with a reddish brown AEC product and in the 2nd sequence an avidin conjugated peroxidase with a blue colour product via 4CN. 2 antigens on the same cell in the same section can be detected by this method. Double labelled cells were coloured violet. The method is particularly suitable for loosely situated cells.