Paper Submission & Registration
8th Dutch Bio-Medical Engineering Conference
Go-previous
15:20   Brain - II
Chair: Raf van Hoof
15:20
15 mins
3D segmentation of nuclei and axons in third harmonic generation images of human brain tissue with U-net
Max Blokker, Pieter Wesseling, Philip de Witt Hamer, Mitko Veta, Marloes Groot
Abstract: Label-free microscopy allows for on-site surgical pathology workflows where tissue can be imaged in the operating theatre with the additional benefit of real-time image analysis. It is our goal to support neurosurgeons during surgery on the issue of tumour boundary detection. Freshly resected glioma tissue can be visualized on a portable setup with third harmonic generation (THG) microscopy in order to reveal structures like cells, nuclei and myelinated axons. Identification and quantification of these structures allows for classification of brain tissue as shown by Z. Zhang et al. (2019), who relied on anisotropic diffusion and active contour to make a distinction between 2D THG normal brain and glioma brain images with 96.6% sensitivity and 95.5% specificity, in nearly perfect (93%) agreement with pathology. This algorithm is however far slower than the THG data acquisition itself. Therefore, an artificial intelligence approach is required. Building upon the work of Zhang, 3D versions of above-mentioned algorithms were iteratively applied and optimized on THG z-stacks in order to generate volumetric ground-truths. Careful post-processing based on feature size, sphericity and ellipsoid ratio allowed for splitting of segmented features into three classes: nuclei, myelinated fibers and inconclusive nuclei/fiber features. The laborious nature of this ground-truth creation process limited the number of data that could be successfully annotated to one glioma patient, illustrating the need for an AI approach. We trained a 3D U-net in semantically segmenting this data outputting a two-channel volume block containing nuclei and myelinated fibers. A custom-built loss function incorporating the generalized Dice loss forced the model to appoint structures in the inconclusive class to one of the two output classes. Testing the trained model on two z-stacks left-out of the training data revealed a structure similarity coefficient between ground-truth and prediction of up to 0.8. Predictions generated on out-of-bounds data like very dense fibers or noisy glioma images proved fruitful yet left room for improvement. It is our goal to further optimize the ground-truth creation process to include more and diverse patient data and ultimately build a 3D segmentation model capable of accurately segmenting nuclei and myelinated fibers in THG brain data.
15:35
15 mins
In vivo magnetic resonance imaging using a custom built 50 mT scanner
Tom O'Reilly, Bart de Vos, Andrew Webb
Abstract: Commercial MRI systems cost millions of euros to purchase, require large shielded spaces to house, are extremely expensive to maintain. These factors together means that their distribution is confined to centrally-located medical centres in large towns and cities. Globally over 70% of the world’s population has absolutely no access to MRI, and clinical conditions which could benefit from even very simple scans cannot be treated. If low-field MRI could be made more portable, accessible and sustainable then it would open up new opportunities in both developed and developing countries. This work describes the design, construction and testing of a low-field MRI system for in vivo imaging of adults and young children. The system consists of a Halbach magnet of twenty three rings, with two layers of N48 neodymium boron iron (NdBFe) magnets (12×12×12 mm3) per ring. The array has a clear bore of 27 cm, and a length of 50 cm. The total weight of the magnet including all components is ~75 kg. The magnet is positioned inside a 62.5×62.5×85 cm Faraday cage constructed from 2 mm thick aluminium sheets. Gradients coils were constructed using 1.5 mm diameter copper wire pressed into 3D printed formers. A triple-axis custom-designed gradient amplifier has been interfaced with the gradient drivers of the MR console. The RF pulses were amplified by a custom built 1 kW RF amplifier. Since scanning is not performed in a shielded room, the volunteer is draped in a conductive cloth to reduce EMI. A 3D Turbo spin echo sequence was used to acquire in vivo images of the adult knee and calf muscle at a resolution of 1×1×3 mm in 10 minutes 40 seconds and of the adult brain at a resolution of 1.85×1.85×3.5 mm in 10 minutes and 8 seconds. Relaxation time mapping gives T1 and T2 values of 160 and 45 ms for muscle, 120 and 110 ms for subcutaneous fat and 250 ms and 110 ms for white/gray matter. The 50 mT system costs approximately 15k euros to construct, and provides a potential solution to the inaccessibility of MRI to much of the world.
15:50
15 mins
Automatic collateral scoring from 3D CTA images
Jiahang Su, Lennard Wolff, Adriaan C.G.M van Es, Wim van Zwam, Charles Majoie, Diederik W.J Dippel, Aad van der Lugt, Wiro .J Niessen, Theo van Walsum
Abstract: Background: Collateral status is relevant for therapeutic decision making for ischemic stroke patients. Patients with good collaterals will mostly likely to have better treatment outcome after EVT. The existing CTA-based collateral scoring system are all visual grading system using coarse definitions. Purpose of our work is to develop and asses an automatic collateral scoring method Method: To this extent, we proposed a three-stage algorithm that consists of pre-processing, deep learning vessel extraction and classification. In the pre-processing, the brain, MCA region and hemispheres are defined by registration to an atlas. In the vessel extraction part, we employ a deep convolutional neural network (CNN), which was trained by 7 fully annotated brains and 20 partially annotated brain cubes. In the classification step, we design features using the ratios of vessel length and volumes in the occluded side and the contralateral side, and developed a multi-class classification model (random forest) that can predict a 4-grade collateral score from the designed features. Results: The method was assessed on 269 images acquired in the clinical routine at 14 intervention centers of Netherlands. Selection criteria for the image data was good image quality, peak arterial phase, equilibrium phase, peak venous phase and a slice thickness less than or equal to 1.5 mm. The collateral score reference standard was the consensus score obtained from 3 radiologists with 10 to 30 years of experience. Interobserver accuracy was 0.6, accuracy of the observer’s vs ground truth was 0.8. The achieved accuracy of the automatic method on the 4-grade collateral score prediction is 0.8, with an average error distance of 0.25 (out of 4). The error is comparable to the interobserver variation, and the results are comparable to the performance of two radiologists with 10 to 30 years of experience. Conclusion: Automated collateral scoring with an accuracy similar to experienced observers is possible.
16:05
15 mins
Wavelet function evaluation for ASD classification in adolescents based on functional MRI
Ramona Cirstian, Antoine Bernas, Svitlana Zinger, A.P. Aldenkamp
Abstract: The autism spectrum disorder (ASD) comprises neurodevelopmental conditions such as autism which is characterized by impaired social abilities, restrictive behaviour and learning disabilities. Due to the lack of an ASD biomarker, no objective diagnosis can be made, and medical professionals rely on observing behaviour for placing a subjective diagnosis. In the last decade, several attempts at developing tools that aid in the objective ASD diagnosis have been made using functional MRI scans in order to evaluate temporal dynamics in different brain networks. Wavelet transform has been proved, in the past, to be a powerful tool in studying the correlation between time-series extracted from fMRI scans. The objective of this study is to examine the performance of multiple wavelet transform types in ASD objective diagnosis by comparing three mother wavelets: derivative of Gaussian (DOG), Morlet and Paul. Two study groups are used in this experiment: the in-house group (12 ASD and 12 control participants) and the Leuven group (12 ASD and 18 control participants). The ’time of in-phase coherence’ is obtained from the wavelet coherence maps of each mother wavelet in order to train the classifiers and evaluate the performance of the diagnosis. The Morlet mother wavelet returned the best performance scores with an up to 87.5% accuracy, followed by Paul (83.3% accuracy) and DOG (80% accuracy). The study concludes that the shape and width of the mother wavelet is very important to the results as well as whether the wavelet function is complex or real valued. Morlet mother wavelet appears to be the best choice for the type of signals used in this study due to the balance between time and frequency resolution.


end %-->