Finn Behrendt

Finn Behrendt, Debayan Bhattacharya, Julia Krüger et al.

The use of supervised deep learning techniques to detect pathologies in brain MRI scans can be challenging due to the diversity of brain anatomy and the need for annotated data sets. An alternative approach is to use unsupervised anomaly detection, which only requires sample-level labels of healthy brains to create a reference representation. This reference representation can then be compared to unhealthy brain anatomy in a pixel-wise manner to identify abnormalities. To accomplish this, generative models are needed to create anatomically consistent MRI scans of healthy brains. While recent diffusion models have shown promise in this task, accurately generating the complex structure of the human brain remains a challenge. In this paper, we propose a method that reformulates the generation task of diffusion models as a patch-based estimation of healthy brain anatomy, using spatial context to guide and improve reconstruction. We evaluate our approach on data of tumors and multiple sclerosis lesions and demonstrate a relative improvement of 25.1% compared to existing baselines.

Debayan Bhattacharya, Benjamin Tobias Becker, Finn Behrendt et al.

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.

Debayan Bhattacharya, Finn Behrendt, Benjamin Tobias Becker et al.

Paranasal anomalies are commonly discovered during routine radiological screenings and can present with a wide range of morphological features. This diversity can make it difficult for convolutional neural networks (CNNs) to accurately classify these anomalies, especially when working with limited datasets. Additionally, current approaches to paranasal anomaly classification are constrained to identifying a single anomaly at a time. These challenges necessitate the need for further research and development in this area. In this study, we investigate the feasibility of using a 3D convolutional neural network (CNN) to classify healthy maxillary sinuses (MS) and MS with polyps or cysts. The task of accurately identifying the relevant MS volume within larger head and neck Magnetic Resonance Imaging (MRI) scans can be difficult, but we develop a straightforward strategy to tackle this challenge. Our end-to-end solution includes the use of a novel sampling technique that not only effectively localizes the relevant MS volume, but also increases the size of the training dataset and improves classification results. Additionally, we employ a multiple instance ensemble prediction method to further boost classification performance. Finally, we identify the optimal size of MS volumes to achieve the highest possible classification performance on our dataset. With our multiple instance ensemble prediction strategy and sampling strategy, our 3D CNNs achieve an F1 of 0.85 whereas without it, they achieve an F1 of 0.70. We demonstrate the feasibility of classifying anomalies in the MS. We propose a data enlarging strategy alongside a novel ensembling strategy that proves to be beneficial for paranasal anomaly classification in the MS.

Finn Behrendt, Debayan Bhattacharya, Julia Krüger et al.

Radiographs are a versatile diagnostic tool for the detection and assessment of pathologies, for treatment planning or for navigation and localization purposes in clinical interventions. However, their interpretation and assessment by radiologists can be tedious and error-prone. Thus, a wide variety of deep learning methods have been proposed to support radiologists interpreting radiographs. Mostly, these approaches rely on convolutional neural networks (CNN) to extract features from images. Especially for the multi-label classification of pathologies on chest radiographs (Chest X-Rays, CXR), CNNs have proven to be well suited. On the Contrary, Vision Transformers (ViTs) have not been applied to this task despite their high classification performance on generic images and interpretable local saliency maps which could add value to clinical interventions. ViTs do not rely on convolutions but on patch-based self-attention and in contrast to CNNs, no prior knowledge of local connectivity is present. While this leads to increased capacity, ViTs typically require an excessive amount of training data which represents a hurdle in the medical domain as high costs are associated with collecting large medical data sets. In this work, we systematically compare the classification performance of ViTs and CNNs for different data set sizes and evaluate more data-efficient ViT variants (DeiT). Our results show that while the performance between ViTs and CNNs is on par with a small benefit for ViTs, DeiTs outperform the former if a reasonably large data set is available for training.

Finn Behrendt, Marcel Bengs, Frederik Rogge et al.

The detection of lesions in magnetic resonance imaging (MRI)-scans of human brains remains challenging, time-consuming and error-prone. Recently, unsupervised anomaly detection (UAD) methods have shown promising results for this task. These methods rely on training data sets that solely contain healthy samples. Compared to supervised approaches, this significantly reduces the need for an extensive amount of labeled training data. However, data labelling remains error-prone. We study how unhealthy samples within the training data affect anomaly detection performance for brain MRI-scans. For our evaluations, we consider three publicly available data sets and use autoencoders (AE) as a well-established baseline method for UAD. We systematically evaluate the effect of impured training data by injecting different quantities of unhealthy samples to our training set of healthy samples from T1-weighted MRI-scans. We evaluate a method to identify falsely labeled samples directly during training based on the reconstruction error of the AE. Our results show that training with impured data decreases the UAD performance notably even with few falsely labeled samples. By performing outlier removal directly during training based on the reconstruction-loss, we demonstrate that falsely labeled data can be detected and removed to mitigate the effect of falsely labeled data. Overall, we highlight the importance of clean data sets for UAD in brain MRI and demonstrate an approach for detecting falsely labeled data directly during training.

Finn Behrendt, Debayan Bhattacharya, Robin Mieling et al.

Unsupervised Anomaly Detection (UAD) methods rely on healthy data distributions to identify anomalies as outliers. In brain MRI, a common approach is reconstruction-based UAD, where generative models reconstruct healthy brain MRIs, and anomalies are detected as deviations between input and reconstruction. However, this method is sensitive to imperfect reconstructions, leading to false positives that impede the segmentation. To address this limitation, we construct multiple reconstructions with probabilistic diffusion models. We then analyze the resulting distribution of these reconstructions using the Mahalanobis distance to identify anomalies as outliers. By leveraging information about normal variations and covariance of individual pixels within this distribution, we effectively refine anomaly scoring, leading to improved segmentation. Our experimental results demonstrate substantial performance improvements across various data sets. Specifically, compared to relying solely on single reconstructions, our approach achieves relative improvements of 15.9%, 35.4%, 48.0%, and 4.7% in terms of AUPRC for the BRATS21, ATLAS, MSLUB and WMH data sets, respectively.

Debayan Bhattacharya, Sarah Latus, Finn Behrendt et al.

Needle positioning is essential for various medical applications such as epidural anaesthesia. Physicians rely on their instincts while navigating the needle in epidural spaces. Thereby, identifying the tissue structures may be helpful to the physician as they can provide additional feedback in the needle insertion process. To this end, we propose a deep neural network that classifies the tissues from the phase and intensity data of complex OCT signals acquired at the needle tip. We investigate the performance of the deep neural network in a limited labelled dataset scenario and propose a novel contrastive pretraining strategy that learns invariant representation for phase and intensity data. We show that with 10% of the training set, our proposed pretraining strategy helps the model achieve an F1 score of 0.84 whereas the model achieves an F1 score of 0.60 without it. Further, we analyse the importance of phase and intensity individually towards tissue classification.

Finn Behrendt, Debayan Bhattacharya, Robin Mieling et al.

The application of supervised models to clinical screening tasks is challenging due to the need for annotated data for each considered pathology. Unsupervised Anomaly Detection (UAD) is an alternative approach that aims to identify any anomaly as an outlier from a healthy training distribution. A prevalent strategy for UAD in brain MRI involves using generative models to learn the reconstruction of healthy brain anatomy for a given input image. As these models should fail to reconstruct unhealthy structures, the reconstruction errors indicate anomalies. However, a significant challenge is to balance the accurate reconstruction of healthy anatomy and the undesired replication of abnormal structures. While diffusion models have shown promising results with detailed and accurate reconstructions, they face challenges in preserving intensity characteristics, resulting in false positives. We propose conditioning the denoising process of diffusion models with additional information derived from a latent representation of the input image. We demonstrate that this conditioning allows for accurate and local adaptation to the general input intensity distribution while avoiding the replication of unhealthy structures. We compare the novel approach to different state-of-the-art methods and for different data sets. Our results show substantial improvements in the segmentation performance, with the Dice score improved by 11.9%, 20.0%, and 44.6%, for the BraTS, ATLAS and MSLUB data sets, respectively, while maintaining competitive performance on the WMH data set. Furthermore, our results indicate effective domain adaptation across different MRI acquisitions and simulated contrasts, an important attribute for general anomaly detection methods. The code for our work is available at https://github.com/FinnBehrendt/Conditioned-Diffusion-Models-UAD

Debayan Bhattacharya, Konrad Reuter, Finn Behrendt et al.

Commonly employed in polyp segmentation, single image UNet architectures lack the temporal insight clinicians gain from video data in diagnosing polyps. To mirror clinical practices more faithfully, our proposed solution, PolypNextLSTM, leverages video-based deep learning, harnessing temporal information for superior segmentation performance with the least parameter overhead, making it possibly suitable for edge devices. PolypNextLSTM employs a UNet-like structure with ConvNext-Tiny as its backbone, strategically omitting the last two layers to reduce parameter overhead. Our temporal fusion module, a Convolutional Long Short Term Memory (ConvLSTM), effectively exploits temporal features. Our primary novelty lies in PolypNextLSTM, which stands out as the leanest in parameters and the fastest model, surpassing the performance of five state-of-the-art image and video-based deep learning models. The evaluation of the SUN-SEG dataset spans easy-to-detect and hard-to-detect polyp scenarios, along with videos containing challenging artefacts like fast motion and occlusion. Comparison against 5 image-based and 5 video-based models demonstrates PolypNextLSTM's superiority, achieving a Dice score of 0.7898 on the hard-to-detect polyp test set, surpassing image-based PraNet (0.7519) and video-based PNSPlusNet (0.7486). Notably, our model excels in videos featuring complex artefacts such as ghosting and occlusion. PolypNextLSTM, integrating pruned ConvNext-Tiny with ConvLSTM for temporal fusion, not only exhibits superior segmentation performance but also maintains the highest frames per speed among evaluated models. Access code here https://github.com/mtec-tuhh/PolypNextLSTM

Debayan Bhattacharya, Finn Behrendt, Benjamin Tobias Becker et al.

Purpose: Paranasal anomalies, frequently identified in routine radiological screenings, exhibit diverse morphological characteristics. Due to the diversity of anomalies, supervised learning methods require large labelled dataset exhibiting diverse anomaly morphology. Self-supervised learning (SSL) can be used to learn representations from unlabelled data. However, there are no SSL methods designed for the downstream task of classifying paranasal anomalies in the maxillary sinus (MS). Methods: Our approach uses a 3D Convolutional Autoencoder (CAE) trained in an unsupervised anomaly detection (UAD) framework. Initially, we train the 3D CAE to reduce reconstruction errors when reconstructing normal maxillary sinus (MS) image. Then, this CAE is applied to an unlabelled dataset to generate coarse anomaly locations by creating residual MS images. Following this, a 3D Convolutional Neural Network (CNN) reconstructs these residual images, which forms our SSL task. Lastly, we fine-tune the encoder part of the 3D CNN on a labelled dataset of normal and anomalous MS images. Results: The proposed SSL technique exhibits superior performance compared to existing generic self-supervised methods, especially in scenarios with limited annotated data. When trained on just 10% of the annotated dataset, our method achieves an Area Under the Precision-Recall Curve (AUPRC) of 0.79 for the downstream classification task. This performance surpasses other methods, with BYOL attaining an AUPRC of 0.75, SimSiam at 0.74, SimCLR at 0.73 and Masked Autoencoding using SparK at 0.75. Conclusion: A self-supervised learning approach that inherently focuses on localizing paranasal anomalies proves to be advantageous, particularly when the subsequent task involves differentiating normal from anomalous maxillary sinuses. Access our code at https://github.com/mtec-tuhh/self-supervised-paranasal-anomaly

Ecem Sogancioglu, Bram van Ginneken, Finn Behrendt et al.

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.

Finn Behrendt, Debayan Bhattacharya, Lennart Maack et al.

Supervised deep learning techniques show promise in medical image analysis. However, they require comprehensive annotated data sets, which poses challenges, particularly for rare diseases. Consequently, unsupervised anomaly detection (UAD) emerges as a viable alternative for pathology segmentation, as only healthy data is required for training. However, recent UAD anomaly scoring functions often focus on intensity only and neglect structural differences, which impedes the segmentation performance. This work investigates the potential of Structural Similarity (SSIM) to bridge this gap. SSIM captures both intensity and structural disparities and can be advantageous over the classical $l1$ error. However, we show that there is more than one optimal kernel size for the SSIM calculation for different pathologies. Therefore, we investigate an adaptive ensembling strategy for various kernel sizes to offer a more pathology-agnostic scoring mechanism. We demonstrate that this ensembling strategy can enhance the performance of DMs and mitigate the sensitivity to different kernel sizes across varying pathologies, highlighting its promise for brain MRI anomaly detection.

Marcel Bengs, Finn Behrendt, Max-Heinrich Laves et al.

Lesion detection in brain Magnetic Resonance Images (MRIs) remains a challenging task. MRIs are typically read and interpreted by domain experts, which is a tedious and time-consuming process. Recently, unsupervised anomaly detection (UAD) in brain MRI with deep learning has shown promising results to provide a quick, initial assessment. So far, these methods only rely on the visual appearance of healthy brain anatomy for anomaly detection. Another biomarker for abnormal brain development is the deviation between the brain age and the chronological age, which is unexplored in combination with UAD. We propose deep learning for UAD in 3D brain MRI considering additional age information. We analyze the value of age information during training, as an additional anomaly score, and systematically study several architecture concepts. Based on our analysis, we propose a novel deep learning approach for UAD with multi-task age prediction. We use clinical T1-weighted MRIs of 1735 healthy subjects and the publicly available BraTs 2019 data set for our study. Our novel approach significantly improves UAD performance with an AUC of 92.60% compared to an AUC-score of 84.37% using previous approaches without age information.

Marcel Bengs, Finn Behrendt, Julia Krüger et al.

Purpose. Brain Magnetic Resonance Images (MRIs) are essential for the diagnosis of neurological diseases. Recently, deep learning methods for unsupervised anomaly detection (UAD) have been proposed for the analysis of brain MRI. These methods rely on healthy brain MRIs and eliminate the requirement of pixel-wise annotated data compared to supervised deep learning. While a wide range of methods for UAD have been proposed, these methods are mostly 2D and only learn from MRI slices, disregarding that brain lesions are inherently 3D and the spatial context of MRI volumes remains unexploited. Methods. We investigate whether using increased spatial context by using MRI volumes combined with spatial erasing leads to improved unsupervised anomaly segmentation performance compared to learning from slices. We evaluate and compare 2D variational autoencoder (VAE) to their 3D counterpart, propose 3D input erasing, and systemically study the impact of the data set size on the performance. Results. Using two publicly available segmentation data sets for evaluation, 3D VAE outperform their 2D counterpart, highlighting the advantage of volumetric context. Also, our 3D erasing methods allow for further performance improvements. Our best performing 3D VAE with input erasing leads to an average DICE score of 31.40% compared to 25.76% for the 2D VAE. Conclusions. We propose 3D deep learning methods for UAD in brain MRI combined with 3D erasing and demonstrate that 3D methods clearly outperform their 2D counterpart for anomaly segmentation. Also, our spatial erasing method allows for further performance improvements and reduces the requirement for large data sets.