Marcel Bengs’s research while affiliated with Hamburg University of Technology and other places

Pulmonary nodules may be an early manifestation of lung cancer, the leading cause of cancer-related deaths among both men and women. Numerous studies have established that deep learning methods can yield high-performance levels in the detection of lung nodules in chest X-rays. However, the lack of gold-standard public datasets slows down the progression of the research and prevents benchmarking of methods for this task. To address this, we organized a public research challenge, NODE21, aimed at the detection and generation of lung nodules in chest X-rays. While the detection track assesses state-of-the-art nodule detection systems, the generation track determines the utility of nodule generation algorithms to augment training data and hence improve the performance of the detection systems. This paper summarizes the results of the NODE21 challenge and performs extensive additional experiments to examine the impact of the synthetically generated nodule training images on the detection algorithm performance.

Lung cancer is a serious disease responsible for millions of deaths every year. Early stages of lung cancer can be manifested in pulmonary lung nodules. To assist radiologists in reducing the number of overseen nodules and to increase the detection accuracy in general, automatic detection algorithms have been proposed. Particularly, deep learning methods are promising. However, obtaining clinically relevant results remains challenging. While a variety of approaches have been proposed for general purpose object detection, these are typically evaluated on benchmark data sets. Achieving competitive performance for specific real-world problems like lung nodule detection typically requires careful analysis of the problem at hand and the selection and tuning of suitable deep learning models. We present a systematic comparison of state-of-the-art object detection algorithms for the task of lung nodule detection. In this regard, we address the critical aspect of class imbalance and and demonstrate a data augmentation approach as well as transfer learning to boost performance. We illustrate how this analysis and a combination of multiple architectures results in state-of-the-art performance for lung nodule detection, which is demonstrated by the proposed model winning the detection track of the Node21 competition. The code for our approach is available at https://github.com/FinnBehrendt/node21-submit.

Motion compensation in radiation therapy is a challenging scenario that requires estimating and forecasting motion of tissue structures to deliver the target dose. Ultrasound offers direct imaging of tissue in real-time and is considered for image guidance in radiation therapy. Recently, fast volumetric ultrasound has gained traction, but motion analysis with such high-dimensional data remains difficult. While deep learning could bring many advantages, such as fast data processing and high performance, it remains unclear how to process sequences of hundreds of image volumes efficiently and effectively. We present a 4D deep learning approach for real-time motion estimation and forecasting using long-term 4D ultrasound data. Using motion traces acquired during radiation therapy combined with various tissue types, our results demonstrate that long-term motion estimation can be performed markerless with a tracking error of $0.35\pm 0.2$ mm and with an inference time of less than 5 ms. Also, we demonstrate forecasting directly from the image data up to 900 ms into the future. Overall, our findings highlight that 4D deep learning is a promising approach for motion analysis during radiotherapy.

Robotic assistance in minimally invasive surgery offers numerous advantages for both patient and surgeon. However, the lack of force feedback in robotic surgery is a major limitation, and accurately estimating tool-tissue interaction forces remains a challenge. Image-based force estimation offers a promising solution without the need to integrate sensors into surgical tools. In this indirect approach, interaction forces are derived from the observed deformation, with learning-based methods improving accuracy and real-time capability. However, the relationship between deformation and force is determined by the stiffness of the tissue. Consequently, both deformation and local tissue properties must be observed for an approach applicable to heterogeneous tissue. In this work, we use optical coherence tomography, which can combine the detection of tissue deformation with shear wave elastography in a single modality. We present a multi-input deep learning network for processing of local elasticity estimates and volumetric image data. Our results demonstrate that accounting for elastic properties is critical for accurate image-based force estimation across different tissue types and properties. Joint processing of local elasticity information yields the best performance throughout our phantom study. Furthermore, we test our approach on soft tissue samples that were not present during training and show that generalization to other tissue properties is possible.

Using deep learning techniques, anomalies in the paranasal sinus system can be detected automatically in MRI images and can be further analyzed and classified based on their volume, shape and other parameters like local contrast. However due to limited training data, traditional supervised learning methods often fail to generalize. Existing deep learning methods in paranasal anomaly classification have been used to diagnose at most one anomaly. In our work, we consider three anomalies. Specifically, we employ a 3D CNN to separate maxillary sinus volumes without anomaly from maxillary sinus volumes with anomaly. To learn robust representations from a small labelled dataset, we propose a novel learning paradigm that combines contrastive loss and cross-entropy loss. Particularly, we use a supervised contrastive loss that encourages embeddings of maxillary sinus volumes with and without anomaly to form two distinct clusters while the cross-entropy loss encourages the 3D CNN to maintain its discriminative ability. We report that optimising with both losses is advantageous over optimising with only one loss. We also find that our training strategy leads to label efficiency. With our method, a 3D CNN classifier achieves an AUROC of 0.85 ± 0.03 while a 3D CNN classifier optimised with cross-entropy loss achieves an AUROC of 0.66 ± 0.1. Our source code is available at https://github.com/dawnofthedebayan/SupConCE_MICCAI_22.

Using deep learning techniques, anomalies in the paranasal sinus system can be detected automatically in MRI images and can be further analyzed and classified based on their volume, shape and other parameters like local contrast. However due to limited training data, traditional supervised learning methods often fail to generalize. Existing deep learning methods in paranasal anomaly classification have been used to diagnose at most one anomaly. In our work, we consider three anomalies. Specifically, we employ a 3D CNN to separate maxillary sinus volumes without anomalies from maxillary sinus volumes with anomalies. To learn robust representations from a small labelled dataset, we propose a novel learning paradigm that combines contrastive loss and cross-entropy loss. Particularly, we use a supervised contrastive loss that encourages embeddings of maxillary sinus volumes with and without anomaly to form two distinct clusters while the cross-entropy loss encourages the 3D CNN to maintain its discriminative ability. We report that optimising with both losses is advantageous over optimising with only one loss. We also find that our training strategy leads to label efficiency. With our method, a 3D CNN classifier achieves an AUROC of 0.85 while a 3D CNN classifier optimised with cross-entropy loss achieves an AUROC of 0.66.

The increasing workloads for radiologists in clinical practice lead to the need for an automatic support tool for anomaly detection in brain MRI-scans. While supervised learning methods can detect and localize lesions in brain MRI-scans, the need for large, balanced data sets with pixel-level annotations limits their use. In contrast, unsupervised anomaly detection (UAD) models only require healthy brain data for training. Despite the inherent 3D structure of brain MRI-scans, most UAD studies focus on slice-wise processing. In this work, we capture the inter-slice dependencies of the human brain using recurrent neural networks (RNN) and transformer-based self-attention mechanisms together with variational autoencoders (VAE). We show that by this we can improve both reconstruction quality and UAD performance while the number of parameters remain similar to the 2D approach where the slices are processed individually.

Objectives Motion compensation is an interesting approach to improve treatments of moving structures. For example, target motion can substantially affect dose delivery in radiation therapy, where methods to detect and mitigate the motion are widely used. Recent advances in fast, volumetric ultrasound have rekindled the interest in ultrasound for motion tracking. We present a setup to evaluate ultrasound based motion tracking and we study the effect of imaging rate and motion artifacts on its performance. Methods We describe an experimental setup to acquire markerless 4D ultrasound data with precise ground truth from a robot and evaluate different real-world trajectories and system settings toward accurate motion estimation. We analyze motion artifacts in continuously acquired data by comparing to data recorded in a step-and-shoot fashion. Furthermore, we investigate the trade-off between the imaging frequency and resolution. Results The mean tracking errors show that continuously acquired data leads to similar results as data acquired in a step-and-shoot fashion. We report mean tracking errors up to 2.01 mm and 1.36 mm on the continuous data for the lower and higher resolution, respectively, while step-and-shoot data leads to mean tracking errors of 2.52 mm and 0.98 mm. Conclusions We perform a quantitative analysis of different system settings for motion tracking with 4D ultrasound. We can show that precise tracking is feasible and additional motion in continuously acquired data does not impair the tracking. Moreover, the analysis of the frequency resolution trade-off shows that a high imaging resolution is beneficial in ultrasound tracking.

Ultrasound shear wave elasticity imaging is a valuable tool for quantifying the elastic properties of tissue. Typically, the shear wave velocity is derived and mapped to an elasticity value, which neglects information such as the shape of the propagating shear wave or push sequence characteristics. We present 3D spatio-temporal CNNs for fast local elasticity estimation from ultrasound data. This approach is based on retrieving elastic properties from shear wave propagation within small local regions. A large training data set is acquired with a robot from homogeneous gelatin phantoms ranging from 17.42 kPa to 126.05 kPa with various push locations. The results show that our approach can estimate elastic properties on a pixelwise basis with a mean absolute error of 5.014.37 kPa. Furthermore, we estimate local elasticity independent of the push location and can even perform accurate estimates inside the push region. For phantoms with embedded inclusions, we report a 53.93% lower MAE (7.50 kPa) and on the background of 85.24% (1.64 kPa) compared to a conventional shear wave method. Overall, our method offers fast local estimations of elastic properties with small spatio-temporal window sizes.