Deep learning-a first meta-survey of selected reviews across scientific disciplines, their commonalities, challenges and research impact

MedShapeNet - a large-scale dataset of 3D medical shapes for computer vision

OBJECTIVES

The shape is commonly used to describe the objects. State-of-the-art algorithms in medical imaging are predominantly diverging from computer vision, where voxel grids, meshes, point clouds, and implicit surface models are used. This is seen from the growing popularity of ShapeNet (51,300 models) and Princeton ModelNet (127,915 models). However, a large collection of anatomical shapes (e.g., bones, organs, vessels) and 3D models of surgical instruments is missing.

METHODS

We present MedShapeNet to translate data-driven vision algorithms to medical applications and to adapt state-of-the-art vision algorithms to medical problems. As a unique feature, we directly model the majority of shapes on the imaging data of real patients. We present use cases in classifying brain tumors, skull reconstructions, multi-class anatomy completion, education, and 3D printing.

RESULTS

By now, MedShapeNet includes 23 datasets with more than 100,000 shapes that are paired with annotations (ground truth). Our data is freely accessible via a web interface and a Python application programming interface and can be used for discriminative, reconstructive, and variational benchmarks as well as various applications in virtual, augmented, or mixed reality, and 3D printing.

CONCLUSIONS

MedShapeNet contains medical shapes from anatomy and surgical instruments and will continue to collect data for benchmarks and applications. The project page is: https://medshapenet.ikim.nrw/.

Analyzing speed-difference impact on freeway joint injury severities of Leading-Following vehicles using statistical and data-driven models

Rear-end (RE) crashes are notably prevalent and pose a substantial risk on freeways. This paper explores the correlation between speed difference among the following and leading vehicles (Δν) and RE crash risk. Three joint models, comprising both uncorrelated and correlated joint random-parameters bivariate probit (RPBP) approaches (statistical methods) and a cross-stitch multilayer perceptron (CS-MLP) network (a data-driven method), were estimated and compared against three separate models: Support Vector Machines (SVM), eXtreme Gradient Boosting (XGBoost), and MLP networks (all data-driven methods). Data on 15,980 two-vehicle RE crashes were collected over a two-year period, from January 1, 2021, to December 31, 2022, considering two possible levels of injury severity: no injury and injury/fatality for both drivers of following and leading vehicles. The comparative performance analysis demonstrates the superior predictive capability of the CS-MLP network over the uncorrelated/correlated joint RPBP model, SVM, XGBoost, and MLP networks in terms of recall, F-1 Score, and AUC. Significantly, numerous shared variables influence the injury severity outcomes for the following and leading vehicles across both statistical and data-driven approaches. Among these factors, the following vehicle (a truck) and the leading vehicle (a passenger car) demonstrate contrasting effects on the injury severity outcomes for both vehicles. Furthermore, the SHapley Additive exPlanations (SHAP) values from the CS-MLP network visually show the relationship between Δν and injury severity, revealing non-linear trends unlike the average effects shown by statistical methods. They indicate that the least injury outcomes for both following and leading vehicles occurs at a Δν of 0 to 10 mph, matching observed patterns in RE crash data. Additionally, a marked variation in the trend of SHAP values for the two vehicles is noted as the speed difference increases. Therefore, the findings affirm the superior performance of joint model development and substantiate the non-linear impacts of speed difference on injury outcomes. The adoption of dynamic speed control measures is recommended to mitigate the injury outcomes involved in two-vehicle RE crashes.

Back to the Roots: Reconstructing Large and Complex Cranial Defects using an Image-based Statistical Shape Model

Designing implants for large and complex cranial defects is a challenging task, even for professional designers. Current efforts on automating the design process focused mainly on convolutional neural networks (CNN), which have produced state-of-the-art results on reconstructing synthetic defects. However, existing CNN-based methods have been difficult to translate to clinical practice in cranioplasty, as their performance on large and complex cranial defects remains unsatisfactory. In this paper, we present a statistical shape model (SSM) built directly on the segmentation masks of the skulls represented as binary voxel occupancy grids and evaluate it on several cranial implant design datasets. Results show that, while CNN-based approaches outperform the SSM on synthetic defects, they are inferior to SSM when it comes to large, complex and real-world defects. Experienced neurosurgeons evaluate the implants generated by the SSM to be feasible for clinical use after minor manual corrections. Datasets and the SSM model are publicly available at https://github.com/Jianningli/ssm .

GAN-based generation of realistic 3D volumetric data: A systematic review and taxonomy

With the massive proliferation of data-driven algorithms, such as deep learning-based approaches, the availability of high-quality data is of great interest. Volumetric data is very important in medicine, as it ranges from disease diagnoses to therapy monitoring. When the dataset is sufficient, models can be trained to help doctors with these tasks. Unfortunately, there are scenarios where large amounts of data is unavailable. For example, rare diseases and privacy issues can lead to restricted data availability. In non-medical fields, the high cost of obtaining enough high-quality data can also be a concern. A solution to these problems can be the generation of realistic synthetic data using Generative Adversarial Networks (GANs). The existence of these mechanisms is a good asset, especially in healthcare, as the data must be of good quality, realistic, and without privacy issues. Therefore, most of the publications on volumetric GANs are within the medical domain. In this review, we provide a summary of works that generate realistic volumetric synthetic data using GANs. We therefore outline GAN-based methods in these areas with common architectures, loss functions and evaluation metrics, including their advantages and disadvantages. We present a novel taxonomy, evaluations, challenges, and research opportunities to provide a holistic overview of the current state of volumetric GANs.

k-strip: A novel segmentation algorithm in k-space for the application of skull stripping

BACKGROUND AND OBJECTIVE

We present a novel deep learning-based skull stripping algorithm for magnetic resonance imaging (MRI) that works directly in the information rich complex valued k-space.

METHODS

Using four datasets from different institutions with a total of around 200,000 MRI slices, we show that our network can perform skull-stripping on the raw data of MRIs while preserving the phase information which no other skull stripping algorithm is able to work with. For two of the datasets, skull stripping performed by HD-BET (Brain Extraction Tool) in the image domain is used as the ground truth, whereas the third and fourth dataset comes with per-hand annotated brain segmentations.

RESULTS

All four datasets were very similar to the ground truth (DICE scores of 92 %-99 % and Hausdorff distances of under 5.5 pixel). Results on slices above the eye-region reach DICE scores of up to 99 %, whereas the accuracy drops in regions around the eyes and below, with partially blurred output. The output of k-Strip often has smoothed edges at the demarcation to the skull. Binary masks are created with an appropriate threshold.

CONCLUSION

With this proof-of-concept study, we were able to show the feasibility of working in the k-space frequency domain, preserving phase information, with consistent results. Besides preserving valuable information for further diagnostics, this approach makes an immediate anonymization of patient data possible, already before being transformed into the image domain. Future research should be dedicated to discovering additional ways the k-space can be used for innovative image analysis and further workflows.

"A net for everyone": fully personalized and unsupervised neural networks trained with longitudinal data from a single patient

BACKGROUND

With the rise in importance of personalized medicine and deep learning, we combine the two to create personalized neural networks. The aim of the study is to show a proof of concept that data from just one patient can be used to train deep neural networks to detect tumor progression in longitudinal datasets.

METHODS

Two datasets with 64 scans from 32 patients with glioblastoma multiforme (GBM) were evaluated in this study. The contrast-enhanced T1w sequences of brain magnetic resonance imaging (MRI) images were used. We trained a neural network for each patient using just two scans from different timepoints to map the difference between the images. The change in tumor volume can be calculated with this map. The neural networks were a form of a Wasserstein-GAN (generative adversarial network), an unsupervised learning architecture. The combination of data augmentation and the network architecture allowed us to skip the co-registration of the images. Furthermore, no additional training data, pre-training of the networks or any (manual) annotations are necessary.

RESULTS

The model achieved an AUC-score of 0.87 for tumor change. We also introduced a modified RANO criteria, for which an accuracy of 66% can be achieved.

CONCLUSIONS

We show a novel approach to deep learning in using data from just one patient to train deep neural networks to monitor tumor change. Using two different datasets to evaluate the results shows the potential to generalize the method.

Seasonal pigment fluctuation in diploid and polyploid Arabidopsis revealed by machine learning-based phenotyping method PlantServation

Long-term field monitoring of leaf pigment content is informative for understanding plant responses to environments distinct from regulated chambers but is impractical by conventional destructive measurements. We developed PlantServation, a method incorporating robust image-acquisition hardware and deep learning-based software that extracts leaf color by detecting plant individuals automatically. As a case study, we applied PlantServation to examine environmental and genotypic effects on the pigment anthocyanin content estimated from leaf color. We processed >4 million images of small individuals of four Arabidopsis species in the field, where the plant shape, color, and background vary over months. Past radiation, coldness, and precipitation significantly affected the anthocyanin content. The synthetic allopolyploid A. kamchatica recapitulated the fluctuations of natural polyploids by integrating diploid responses. The data support a long-standing hypothesis stating that allopolyploids can inherit and combine the traits of progenitors. PlantServation facilitates the study of plant responses to complex environments termed "in natura".

A deep learning-based interpretable decision tool for predicting high risk of chemotherapy-induced nausea and vomiting in cancer patients prescribed highly emetogenic chemotherapy

OBJECTIVE

This study aims to develop a risk prediction model for chemotherapy-induced nausea and vomiting (CINV) in cancer patients receiving highly emetogenic chemotherapy (HEC) and identify the variables that have the most significant impact on prediction.

METHODS

Data from Tianjin Medical University General Hospital were collected and subjected to stepwise data preprocessing. Deep learning algorithms, including deep forest, and typical machine learning algorithms such as support vector machine (SVM), categorical boosting (CatBoost), random forest, decision tree, and neural network were used to develop the prediction model. After training the model and conducting hyperparameter optimization (HPO) through cross-validation in the training set, the performance was evaluated using the test set. Shapley additive explanations (SHAP), partial dependence plot (PDP), and Local Interpretable Model-Agnostic Explanations (LIME) techniques were employed to explain the optimal model. Model performance was assessed using AUC, F1 score, accuracy, specificity, sensitivity, and Brier score.

RESULTS

The deep forest model exhibited good discrimination, outperforming typical machine learning models, with an AUC of 0.850 (95%CI, 0.780-0.919), an F1 score of 0.757, an accuracy of 0.852, a specificity of 0.863, a sensitivity of 0.784, and a Brier score of 0.082. The top five important features in the model were creatinine clearance (Ccr), age, gender, anticipatory nausea and vomiting, and antiemetic regimen. Among these, Ccr had the most significant predictive value. The risk of CINV decreased with increased Ccr and age, while it was higher in the presence of anticipatory nausea and vomiting, female gender, and non-standard antiemetic regimen.

CONCLUSION

The deep forest model demonstrated good discrimination in predicting the risk of CINV in cancer patients prescribed HEC. Kidney function, as represented by Ccr, played a crucial role in the model's prediction. The clinical application of this predictive tool can help assess individual risks and improve patient care by proactively optimizing the use of antiemetics in cancer patients receiving HEC.

Medical deep learning-A systematic meta-review

Deep learning has remarkably impacted several different scientific disciplines over the last few years. For example, in image processing and analysis, deep learning algorithms were able to outperform other cutting-edge methods. Additionally, deep learning has delivered state-of-the-art results in tasks like autonomous driving, outclassing previous attempts. There are even instances where deep learning outperformed humans, for example with object recognition and gaming. Deep learning is also showing vast potential in the medical domain. With the collection of large quantities of patient records and data, and a trend towards personalized treatments, there is a great need for automated and reliable processing and analysis of health information. Patient data is not only collected in clinical centers, like hospitals and private practices, but also by mobile healthcare apps or online websites. The abundance of collected patient data and the recent growth in the deep learning field has resulted in a large increase in research efforts. In Q2/2020, the search engine PubMed returned already over 11,000 results for the search term 'deep learning', and around 90% of these publications are from the last three years. However, even though PubMed represents the largest search engine in the medical field, it does not cover all medical-related publications. Hence, a complete overview of the field of 'medical deep learning' is almost impossible to obtain and acquiring a full overview of medical sub-fields is becoming increasingly more difficult. Nevertheless, several review and survey articles about medical deep learning have been published within the last few years. They focus, in general, on specific medical scenarios, like the analysis of medical images containing specific pathologies. With these surveys as a foundation, the aim of this article is to provide the first high-level, systematic meta-review of medical deep learning surveys.

Effect of neural network structure in accelerating performance and accuracy of a convolutional neural network with GPU/TPU for image analytics

BACKGROUND

In deep learning the most significant breakthrough in the field of image recognition, object detection language processing was done by Convolutional Neural Network (CNN). Rapid growth in data and neural networks the performance of the DNN algorithms depends on the computation power and the storage capacity of the devices.

METHODS

In this paper, the convolutional neural network used for various image applications was studied and its acceleration in the various platforms like CPU, GPU, TPU was done. The neural network structure and the computing power and characteristics of the GPU, TPU was analyzed and summarized, the effect of these on accelerating the tasks is also explained. Cross-platform comparison of the CNN was done using three image applications the face mask detection (object detection/Computer Vision), Virus Detection in Plants (Image Classification: agriculture sector), and Pneumonia detection from X-ray Images (Image Classification/medical field).

RESULTS

The CNN implementation was done and a comprehensive comparison was done on the platforms to identify the performance, throughput, bottlenecks, and training time. The CNN layer-wise execution in GPU and TPU is explained with layer-wise analysis. The impact of the fully connected layer and convolutional layer on the network is analyzed. The challenges faced during the acceleration process were discussed and future works are identified.

AVT: Multicenter aortic vessel tree CTA dataset collection with ground truth segmentation masks

In this article, we present a multicenter aortic vessel tree database collection, containing 56 aortas and their branches. The datasets have been acquired with computed tomography angiography (CTA) scans and each scan covers the ascending aorta, the aortic arch and its branches into the head/neck area, the thoracic aorta, the abdominal aorta and the lower abdominal aorta with the iliac arteries branching into the legs. For each scan, the collection provides a semi-automatically generated segmentation mask of the aortic vessel tree (ground truth). The scans come from three different collections and various hospitals, having various resolutions, which enables studying the geometry/shape variabilities of human aortas and its branches from different geographic locations. Furthermore, creating a robust statistical model of the shape of human aortic vessel trees, which can be used for various tasks such as the development of fully-automatic segmentation algorithms for new, unseen aortic vessel tree cases, e.g. by training deep learning-based approaches. Hence, the collection can serve as an evaluation set for automatic aortic vessel tree segmentation algorithms.