Logo: to the web site of Uppsala University

While deep learning has seen consistent improvements in medical imaging applications, one of the biggest hurdles for further gains is the need for large amounts of data. This data is not always readily available for different reasons, e.g., due to a lack of experts to annotate the samples. Equivariant neural networks have proven to be a way to increase data efficiency by, e.g., forgoing the need for data augmentation. This study extends previous research on group-equivariant networks applied to transmission electron microscopy (TEM) images of different types of viruses. It is shown that group equivariant networks, when compared to baseline convolutional networks, obtain higher accuracies by 1.4 to 2.3% when increasing the training epochs from 300 to 2400. Furthermore, data augmentation strategies by rotations and reflections, as well as pre-training on ImageNet, can be skipped. When turning off batch normalization, the performance of the equivariant networks drops about 25% while the baseline fails to converge, implying that equivariance and batch normalization extract improved information from data by different mechanisms. The VGG16 architecture outperforms a smaller custom architecture by 3.9 to 4.4%, while the choice of symmetry group does not impact performance significantly. This study contributes to predictable scaling, which is already of great importance in, e.g., the training of large language models. Predictable scaling is expected to increase in significance in the field of biomedical image analysis since the model and dataset sizes continue to grow.

Transmission electron microscopy (TEM) is an imaging technique used to visualize and analyze nano-sized structures and objects such as virus particles. Light microscopy can be used to diagnose diseases or characterize e.g. blood cells. Since samples under microscopes exhibit certain symmetries, such as global rotation invariance, equivariant neural networks are presumed to be useful. In this study, a baseline convolutional neural network is constructed in the form of the commonly used VGG16 classifier. Thereafter, it is modified to be equivariant to the p4 symmetry group of rotations of multiples of 90 degrees using group convolutions. This yields a number of benefits on a TEM virus dataset, including higher top validation set accuracy by on average 7.6% and faster convergence during training by on average 23.1% of that of the baseline. Similarly, when training and testing on images of blood cells, the convergence time for the equivariant neural network is 7.9% of that of the baseline. From this it is concluded that augmentation strategies for rotation can be skipped. Furthermore, when modelling the accuracy versus amount of TEM virus training data with a power law, the equivariant network has a slope of - 0.43 compared to - 0.26 of the baseline. Thus the equivariant network learns faster than the baseline when more training data is added. This study extends previous research on equivariant neural networks applied to images which exhibit symmetries to isometric transformations.

Objectives: To evaluate an in situ reference acquisition method for implant positions in complete edentulous maxillae using an industrial scanner and allowing for in vivo trueness analysis of the restorative workflow. To assess in vivo trueness and precision of intraoral scanners (IOS) using different acquisition protocols. Furthermore, to compare IOS trueness with impression-based models and implant-supported fixed dentures (IFD) in a parallel study on the same cohort using the same in situ reference scan.

Methods: Six scan-bodies mounted to maxillary implants in five subjects were reference scanned (REF) using an industrial scanner. Subjects were scanned with IOS three times using three different protocols: control (CT), dental floss assisted (DF), and acrylic splint (SP). CAD-files of scan-bodies with inter-aligned analogues were geometry-aligned to REF, and SP. Scan-bodies were aligned to CT and DF in proprietary dental laboratory software and exported with analogue positions. Resulting six CAD-analogues per scan were Globally Aligned using a consistent geometry-based alignment. Deviations were computed after a Reference Point System Alignment at the implant/prosthetic platform for Cartesian axes with a linear Resultant.

Results: Resultant trueness was CT: 41±11 μm, DF: 49±22 μm, SP: 55±8 μm. Resultant precision was CT: 48±7μm, DF: 50±7 μm, SP: 45±6 μm

Conclusions: This method is applicable for assessing trueness of maxillary full-arch implant scans in vivo. The CTprotocol was most accurate. CT trueness showed no difference to digitised impression-based models in parallelstudy. CT was more accurate than IFD in a parallel study. CT displayed similar numerical trueness as existing invitro studies.

Critical significance: Using IOS to acquire full-arch implant scans is controversial. The modified protocol in thispilot shows promising results in the maxilla where great care was taken to manage non-attached tissues when amodified scanning pattern was used. However, other IOS may show varying results in vivo. A completed scan doesnot necessarily equate to an accurate scan.

Objectives: To evaluate a method for in situ reference acquisition of implant positions in complete edentulous maxillae using an industrial scanner. To assess in vivo trueness of full-arch implant-supported fixed dentures (IFD) and dental models based on conventional impressions.

Methods: In five subjects, scan-bodies were mounted to six maxillary implants and scanned three times using an industrial scanner (REF). Original impression-based models used to manufacture existing IFDs, (MOD1), and models fabricated from new polyether impressions, (MOD2), were scanned three times with a laboratory scanner. Scan-bodies were aligned and exported with analogue positions corresponding to implant positions. Implant analogues were mounted onto existing IFDs and scanned three times (BRIDGE). CAD files of scan-bodies with inter-aligned CAD-analogues were geometry-aligned to REF. CAD-analogues were aligned to exported files of MOD1 and MOD2, and to BRIDGE. Resulting six CAD-analogues were Globally Aligned using a consistent geometry-based alignment. Deviations were computed after a Reference Point System Alignment at the implant/prosthetic platform for Cartesian axes and a linear Resultant.

Results:REF precision was 9.3 ± 1 µm. In vivo trueness for Resultant was MOD1: 36±16 µm, MOD2: 28±7 µm and BRIDGE: 70±23 µm, where MOD1 and MOD2 were statistically significantly different from BRIDGE. In vitro manufacturing trueness of Resultant when MOD1 acted reference for BRIDGE was: 69 ± 22.

Conclusions: This method can be applied for assessing in vivo trueness. CAD/CAM processed IFD showed deviations twice that of impression-based models, however, errors from impressions and subsequent model scans were not additive to the entire workflow.

The management of complex mandible fractures, i.e severely comminuted or fractures of edentulous/atrophic mandibles, can be challenging. This is due to the three-dimensional loss of bone, which limits the possibility for accurate anatomic reduction. Virtual surgery planning (VSP) can provide improved accuracy and shorter operating times, but is often not employed for trauma cases because of time constraints and complex user interfaces limited to two-dimensional interaction with three-dimensional data. In this study, we evaluate the accuracy, precision, and time efficiency of the Haptic Assisted Surgery Planning system (HASP), an in-house VSP system that supports stereo graphics, six degrees-of-freedom input and haptics, to improve the surgical planning. Three operators performed planning in HASP on Computed Tomography (CT) and Come Beam Computed Tomography (CBCT) images of a plastic skull model and on twelve retrospective cases with complex mandible fractures. The result shows an accuracy and reproducibility of less than 2mm when using HASP, with an average planning time of 15 minutes, including time for segmentation in the software BoneSplit. This study presents an in-house haptic assisted planning tool for cranio-maxillofacial surgery with high usability that can be used for preoperative planning and evaluation of complex mandible fractures. 

We present how replacing convolutional neural networks with a rotation-equivariant counterpart can be used to reduce the amount of training images needed for classification of whether a cell is cancerous or not. Our hypothesis is that data augmentation schemes by rotation can be replaced, thereby increasing weight sharing and reducing overfitting. The dataset at hand consists of single cell images. We have balanced a subset of almost 9.000 images from healthy patients and patients diagnosed with cancer. Results show that classification accuracy is improved and overfitting reduced if compared to an ordinary convolutional neural network. The results are encouraging and thereby an advancing step towards making screening of patients widely used for the application of oral cancer.

Real‐time virtual reality requires efficient rendering methods to deal with high‐ resolution stereoscopic displays and low latency head‐tracking. Our proposed RayCaching method renders isosurfaces of large volume datasets by amortizing raycasting over several frames and caching primary rays as small bricks that can be efficiently rasterized. An occupancy map in form of a clipmap provides level of detail and ensures that only bricks corresponding to visible points on the isosurface are being cached and rendered. Hard shadows and ambient occlusion from secondary rays are also accumulated and stored in the cache. Our method supports real‐time isosurface rendering with dynamic isovalue and allows stereoscopic visualization and exploration of large volume datasets at framerates suitable for virtual reality applications.