Improving multiple sclerosis lesion segmentation across clinical sites: A federated learning approach with noise-resilient training

Accurately measuring the evolution of Multiple Sclerosis (MS) with magnetic resonance imaging (MRI) critically informs understanding of disease progression and helps to direct therapeutic strategy. Deep learning models have shown promise for automatically segmenting MS lesions, but the scarcity of accurately annotated data hinders progress in this area. Obtaining sufficient data from a single clinical site is challenging and does not address the heterogeneous need for model robustness. Conversely, the collection of data from multiple sites introduces data privacy concerns and potential label noise due to varying annotation standards. To address this dilemma, we explore the use of the federated learning framework while considering label noise. Our approach enables collaboration among multiple clinical sites without compromising data privacy under a federated learning paradigm that incorporates a noise-robust training strategy based on label correction. Specifically, we introduce a Decoupled Hard Label Correction (DHLC) strategy that considers the imbalanced distribution and fuzzy boundaries of MS lesions, enabling the correction of false annotations based on prediction confidence. We also introduce a Centrally Enhanced Label Correction (CELC) strategy, which leverages the aggregated central model as a correction teacher for all sites, enhancing the reliability of the correction process. Extensive experiments conducted on two multi-site datasets demonstrate the effectiveness and robustness of our proposed methods, indicating their potential for clinical applications in multi-site collaborations to train better deep learning models with lower cost in data collection and annotation.

A survey of label-noise deep learning for medical image analysis

Several factors are associated with the success of deep learning. One of the most important reasons is the availability of large-scale datasets with clean annotations. However, obtaining datasets with accurate labels in the medical imaging domain is challenging. The reliability and consistency of medical labeling are some of these issues, and low-quality annotations with label noise usually exist. Because noisy labels reduce the generalization performance of deep neural networks, learning with noisy labels is becoming an essential task in medical image analysis. Literature on this topic has expanded in terms of volume and scope. However, no recent surveys have collected and organized this knowledge, impeding the ability of researchers and practitioners to utilize it. In this work, we presented an up-to-date survey of label-noise learning for medical image domain. We reviewed extensive literature, illustrated some typical methods, and showed unified taxonomies in terms of methodological differences. Subsequently, we conducted the methodological comparison and demonstrated the corresponding advantages and disadvantages. Finally, we discussed new research directions based on the characteristics of medical images. Our survey aims to provide researchers and practitioners with a solid understanding of existing medical label-noise learning, such as the main algorithms developed over the past few years, which could help them investigate new methods to combat with the negative effects of label noise.

LensePro: label noise-tolerant prototype-based network for improving cancer detection in prostate ultrasound with limited annotations

PURPOSE

The standard of care for prostate cancer (PCa) diagnosis is the histopathological analysis of tissue samples obtained via transrectal ultrasound (TRUS) guided biopsy. Models built with deep neural networks (DNNs) hold the potential for direct PCa detection from TRUS, which allows targeted biopsy and subsequently enhances outcomes. Yet, there are ongoing challenges with training robust models, stemming from issues such as noisy labels, out-of-distribution (OOD) data, and limited labeled data.

METHODS

This study presents LensePro, a unified method that not only excels in label efficiency but also demonstrates robustness against label noise and OOD data. LensePro comprises two key stages: first, self-supervised learning to extract high-quality feature representations from abundant unlabeled TRUS data and, second, label noise-tolerant prototype-based learning to classify the extracted features.

RESULTS

Using data from 124 patients who underwent systematic prostate biopsy, LensePro achieves an AUROC, sensitivity, and specificity of 77.9%, 85.9%, and 57.5%, respectively, for detecting PCa in ultrasound. Our model shows it is effective for detecting OOD data in test time, critical for clinical deployment. Ablation studies demonstrate that each component of our method improves PCa detection by addressing one of the three challenges, reinforcing the benefits of a unified approach.

CONCLUSION

Through comprehensive experiments, LensePro demonstrates its state-of-the-art performance for TRUS-based PCa detection. Although further research is necessary to confirm its clinical applicability, LensePro marks a notable advancement in enhancing automated computer-aided systems for detecting prostate cancer in ultrasound.

Contrastive learning of graphs under label noise

In the domain of graph-structured data learning, semi-supervised node classification serves as a critical task, relying mainly on the information from unlabeled nodes and a minor fraction of labeled nodes for training. However, real-world graph-structured data often suffer from label noise, which significantly undermines the performance of Graph Neural Networks (GNNs). This problem becomes increasingly severe in situations where labels are scarce. To tackle this issue of sparse and noisy labels, we propose a novel approach Contrastive Robust Graph Neural Network (CR-GNN), Firstly, considering label sparsity and noise, we employ unsupervised contrastive loss and further incorporate homophily in the graph structure, thus introducing neighbor contrastive loss. Moreover, data augmentation is typically used to construct positive and negative samples in contrastive learning, which may result in inconsistent prediction outcomes. Based on this, we propose a dynamic cross-entropy loss, which selects the nodes with consistent predictions as reliable nodes for cross-entropy loss and benefits to mitigate the overfitting to labeling noise. Finally, we propose cross-space consistency to narrow the semantic gap between the contrast and classification spaces. Extensive experiments on multiple publicly available datasets demonstrate that CR-GNN notably outperforms existing methods in resisting label noise.

Learning from crowds for automated histopathological image segmentation

Automated semantic segmentation of histopathological images is an essential task in Computational Pathology (CPATH). The main limitation of Deep Learning (DL) to address this task is the scarcity of expert annotations. Crowdsourcing (CR) has emerged as a promising solution to reduce the individual (expert) annotation cost by distributing the labeling effort among a group of (non-expert) annotators. Extracting knowledge in this scenario is challenging, as it involves noisy annotations. Jointly learning the underlying (expert) segmentation and the annotators' expertise is currently a commonly used approach. Unfortunately, this approach is frequently carried out by learning a different neural network for each annotator, which scales poorly when the number of annotators grows. For this reason, this strategy cannot be easily applied to real-world CPATH segmentation. This paper proposes a new family of methods for CR segmentation of histopathological images. Our approach consists of two coupled networks: a segmentation network (for learning the expert segmentation) and an annotator network (for learning the annotators' expertise). We propose to estimate the annotators' behavior with only one network that receives the annotator ID as input, achieving scalability on the number of annotators. Our family is composed of three different models for the annotator network. Within this family, we propose a novel modeling of the annotator network in the CR segmentation literature, which considers the global features of the image. We validate our methods on a real-world dataset of Triple Negative Breast Cancer images labeled by several medical students. Our new CR modeling achieves a Dice coefficient of 0.7827, outperforming the well-known STAPLE (0.7039) and being competitive with the supervised method with expert labels (0.7723). The code is available at https://github.com/wizmik12/CRowd_Seg.

Ensembling noisy segmentation masks of blurred sperm images

BACKGROUND

Sperm tail morphology and motility have been demonstrated to be important factors in determining sperm quality for in vitro fertilization. However, many existing computer-aided sperm analysis systems leave the sperm tail out of the analysis, as detecting a few tail pixels is challenging. Moreover, some publicly available datasets for classifying morphological defects contain images limited only to the sperm head. This study focuses on the segmentation of full sperm, which consists of the head and tail parts, and appear alone and in groups.

METHODS

We re-purpose the Feature Pyramid Network to ensemble an input image with multiple masks from state-of-the-art segmentation algorithms using a scale-specific cross-attention module. We normalize homogeneous backgrounds for improved training. The low field depth of microscopes blurs the images, easily confusing human raters in discerning minuscule sperm from large backgrounds. We thus propose evaluation protocols for scoring segmentation models trained on imbalanced data and noisy ground truth.

RESULTS

The neural ensembling of noisy segmentation masks outperforms all single, state-of-the-art segmentation algorithms in full sperm segmentation. Human raters agree more on the head than tail masks. The algorithms also segment the head better than the tail.

CONCLUSIONS

The extensive evaluation of state-of-the-art segmentation algorithms shows that full sperm segmentation is challenging. We release the SegSperm dataset of images from Intracytoplasmic Sperm Injection procedures to spur further progress on full sperm segmentation with noisy and imbalanced ground truth. The dataset is publicly available at https://doi.org/10.34808/6wm7-1159.

A fundus image classification framework for learning with noisy labels

Fundus images are widely used in the screening and diagnosis of eye diseases. Current classification algorithms for computer-aided diagnosis in fundus images rely on large amounts of data with reliable labels. However, the appearance of noisy labels degrades the performance of data-dependent algorithms, such as supervised deep learning. A noisy label learning framework suitable for the multiclass classification of fundus diseases is presented in this paper, which combines data cleansing (DC), adaptive negative learning (ANL), and sharpness-aware minimization (SAM) modules. Firstly, the DC module filters the noisy labels in the training dataset based on the prediction confidence. Then, the ANL module modifies the loss function by choosing complementary labels, which are neither the given labels nor the labels with the highest confidence. Moreover, for better generalization, the SAM module is applied by simultaneously optimizing the loss and its sharpness. Extensive experiments on both private and public datasets show that our method greatly promotes the performance for classification of multiple fundus diseases with noisy labels.

Learning to segment fetal brain tissue from noisy annotations

Automatic fetal brain tissue segmentation can enhance the quantitative assessment of brain development at this critical stage. Deep learning methods represent the state of the art in medical image segmentation and have also achieved impressive results in brain segmentation. However, effective training of a deep learning model to perform this task requires a large number of training images to represent the rapid development of the transient fetal brain structures. On the other hand, manual multi-label segmentation of a large number of 3D images is prohibitive. To address this challenge, we segmented 272 training images, covering 19-39 gestational weeks, using an automatic multi-atlas segmentation strategy based on deformable registration and probabilistic atlas fusion, and manually corrected large errors in those segmentations. Since this process generated a large training dataset with noisy segmentations, we developed a novel label smoothing procedure and a loss function to train a deep learning model with smoothed noisy segmentations. Our proposed methods properly account for the uncertainty in tissue boundaries. We evaluated our method on 23 manually-segmented test images of a separate set of fetuses. Results show that our method achieves an average Dice similarity coefficient of 0.893 and 0.916 for the transient structures of younger and older fetuses, respectively. Our method generated results that were significantly more accurate than several state-of-the-art methods including nnU-Net that achieved the closest results to our method. Our trained model can serve as a valuable tool to enhance the accuracy and reproducibility of fetal brain analysis in MRI.

Learning to segment subcortical structures from noisy annotations with a novel uncertainty-reliability aware learning framework

Accurate segmentation of subcortical structures is an important task in quantitative brain image analysis. Convolutional neural networks (CNNs) have achieved remarkable results in medical image segmentation. However, due to the difficulty of acquiring high-quality annotations of brain subcortical structures, learning segmentation networks using noisy annotations is an inevitable topic. A common practice is to select images or pixels with reliable annotations for training, which usually may not make full use of the information from the training samples, thus affecting the performance of the learned segmentation model. To address the above problem, in this work, we propose a novel robust learning method and denote it as uncertainty-reliability awareness learning (URAL), which can make sufficient use of all training pixels. At each training iteration, the proposed method first selects training pixels with reliable annotations from the set of pixels with uncertain network prediction, by utilizing a small clean validation set following a meta-learning paradigm. Meanwhile, we propose the online prototypical soft label correction (PSLC) method to estimate the pseudo-labels of label-unreliable pixels. Then, the segmentation loss of label-reliable pixels and the semi-supervised segmentation loss of label-unreliable pixels are used to calibrate the total segmentation loss. Finally, we propose a category-wise contrastive regularization to learn compact feature representations of all uncertain training pixels. Comprehensive experiments are performed on two publicly available brain MRI datasets. The proposed method achieves the best Dice scores and MHD values on both datasets compared to several recent state-of-the-art methods under all label noise settings. Our code is available at https://github.com/neulxlx/URAL.

A deep learning framework to classify breast density with noisy labels regularization

BACKGROUND AND OBJECTIVE

Breast density assessed from digital mammograms is a biomarker for higher risk of developing breast cancer. Experienced radiologists assess breast density using the Breast Image and Data System (BI-RADS) categories. Supervised learning algorithms have been developed with this objective in mind, however, the performance of these algorithms depends on the quality of the ground-truth information which is usually labeled by expert readers. These labels are noisy approximations of the ground truth, as there is often intra- and inter-reader variability among labels. Thus, it is crucial to provide a reliable method to obtain digital mammograms matching BI-RADS categories. This paper presents RegL (Labels Regularizer), a methodology that includes different image pre-processes to allow both a correct breast segmentation and the enhancement of image quality through an intensity adjustment, thus allowing the use of deep learning to classify the mammograms into BI-RADS categories. The Confusion Matrix (CM) - CNN network used implements an architecture that models each radiologist's noisy label. The final methodology pipeline was determined after comparing the performance of image pre-processes combined with different DL architectures.

METHODS

A multi-center study composed of 1395 women whose mammograms were classified into the four BI-RADS categories by three experienced radiologists is presented. A total of 892 mammograms were used as the training corpus, 224 formed the validation corpus, and 279 the test corpus.

RESULTS

The combination of five networks implementing the RegL methodology achieved the best results among all the models in the test set. The ensemble model obtained an accuracy of (0.85) and a kappa index of 0.71.

CONCLUSIONS

The proposed methodology has a similar performance to the experienced radiologists in the classification of digital mammograms into BI-RADS categories. This suggests that the pre-processing steps and modelling of each radiologist's label allows for a better estimation of the unknown ground truth labels.

Coarse label refinement for improving prostate cancer detection in ultrasound imaging

PURPOSE

Ultrasound-guided biopsy plays a major role in prostate cancer (PCa) detection, yet is limited by a high rate of false negatives and low diagnostic yield of the current systematic, non-targeted approaches. Developing machine learning models for accurately identifying cancerous tissue in ultrasound would help sample tissues from regions with higher cancer likelihood. A plausible approach for this purpose is to use individual ultrasound signals corresponding to a core as inputs and consider the histopathology diagnosis for the entire core as labels. However, this introduces significant amount of label noise to training and degrades the classification performance. Previously, we suggested that histopathology-reported cancer involvement can be a reasonable approximation for the label noise.

METHODS

Here, we propose an involvement-based label refinement (iLR) method to correct corrupted labels and improve cancer classification. The difference between predicted and true cancer involvements is used to guide the label refinement process. We further incorporate iLR into state-of-the-art methods for learning with noisy labels and predicting cancer involvement.

RESULTS

We use 258 biopsy cores from 70 patients and demonstrate that our proposed label refinement method improves the performance of multiple noise-tolerant approaches and achieves a balanced accuracy, correlation coefficient, and mean absolute error of 76.7%, 0.68, and 12.4, respectively.

CONCLUSIONS

Our key contribution is to leverage a data-centric method to deal with noisy labels using histopathology reports, and improve the performance of prostate cancer diagnosis through a hierarchical training process with label refinement.

Curriculum learning for improved femur fracture classification: Scheduling data with prior knowledge and uncertainty

An adequate classification of proximal femur fractures from X-ray images is crucial for the treatment choice and the patients' clinical outcome. We rely on the commonly used AO system, which describes a hierarchical knowledge tree classifying the images into types and subtypes according to the fracture's location and complexity. In this paper, we propose a method for the automatic classification of proximal femur fractures into 3 and 7 AO classes based on a Convolutional Neural Network (CNN). As it is known, CNNs need large and representative datasets with reliable labels, which are hard to collect for the application at hand. In this paper, we design a curriculum learning (CL) approach that improves over the basic CNNs performance under such conditions. Our novel formulation reunites three curriculum strategies: individually weighting training samples, reordering the training set, and sampling subsets of data. The core of these strategies is a scoring function ranking the training samples. We define two novel scoring functions: one from domain-specific prior knowledge and an original self-paced uncertainty score. We perform experiments on a clinical dataset of proximal femur radiographs. The curriculum improves proximal femur fracture classification up to the performance of experienced trauma surgeons. The best curriculum method reorders the training set based on prior knowledge resulting into a classification improvement of 15%. Using the publicly available MNIST dataset, we further discuss and demonstrate the benefits of our unified CL formulation for three controlled and challenging digit recognition scenarios: with limited amounts of data, under class-imbalance, and in the presence of label noise. The code of our work is available at: https://github.com/ameliajimenez/curriculum-learning-prior-uncertainty.

S-CUDA: Self-cleansing unsupervised domain adaptation for medical image segmentation

Medical image segmentation tasks hitherto have achieved excellent progresses with large-scale datasets, which empowers us to train potent deep convolutional neural networks (DCNNs). However, labeling such large-scale datasets is laborious and error-prone, which leads the noisy (or incorrect) labels to be an ubiquitous problem in the real-world scenarios. In addition, data collected from different sites usually exhibit significant data distribution shift (or domain shift). As a result, noisy label and domain shift become two common problems in medical imaging application scenarios, especially in medical image segmentation, which degrade the performance of deep learning models significantly. In this paper, we identify a novel problem hidden in medical image segmentation, which is unsupervised domain adaptation on noisy labeled data, and propose a novel algorithm named "Self-Cleansing Unsupervised Domain Adaptation" (S-CDUA) to address such issue. S-CUDA sets up a realistic scenario to solve the above problems simultaneously where training data (i.e., source domain) not only shows domain shift w.r.t. unsupervised test data (i.e., target domain) but also contains noisy labels. The key idea of S-CUDA is to learn noise-excluding and domain invariant knowledge from noisy supervised data, which will be applied on the highly corrupted data for label cleansing and further data-recycling, as well as on the test data with domain shift for supervised propagation. To this end, we propose a novel framework leveraging noisy-label learning and domain adaptation techniques to cleanse the noisy labels and learn from trustable clean samples, thus enabling robust adaptation and prediction on the target domain. Specifically, we train two peer adversarial networks to identify high-confidence clean data and exchange them in companions to eliminate the error accumulation problem and narrow the domain gap simultaneously. In the meantime, the high-confidence noisy data are detected and cleansed in order to reuse the contaminated training data. Therefore, our proposed method can not only cleanse the noisy labels in the training set but also take full advantage of the existing noisy data to update the parameters of the network. For evaluation, we conduct experiments on two popular datasets (REFUGE and Drishti-GS) for optic disc (OD) and optic cup (OC) segmentation, and on another public multi-vendor dataset for spinal cord gray matter (SCGM) segmentation. Experimental results show that our proposed method can cleanse noisy labels efficiently and obtain a model with better generalization performance at the same time, which outperforms previous state-of-the-art methods by large margin. Our code can be found at https://github.com/zzdxjtu/S-cuda.

P-DIFF+: Improving learning classifier with noisy labels by Noisy Negative Learning loss

Learning deep neural network (DNN) classifier with noisy labels is a challenging task because the DNN can easily over-fit on these noisy labels due to its high capability. In this paper, we present a very simple but effective training paradigm called P-DIFF+, which can train DNN classifiers but obviously alleviate the adverse impact of noisy labels. Our proposed probability difference distribution implicitly reflects the probability of a training sample to be clean, then this probability is employed to re-weight the corresponding sample during the training process. Moreover, Noisy Negative Learning(NNL) loss can be further employed to re-weight samples. P-DIFF+ can achieve good performance even without prior-knowledge on the noise rate of training samples. Experiments on benchmark datasets demonstrate that P-DIFF+ is superior to the state-of-the-art sample selection methods.

How to handle noisy labels for robust learning from uncertainty

Most deep neural networks (DNNs) are trained with large amounts of noisy labels when they are applied. As DNNs have the high capacity to fit any noisy labels, it is known to be difficult to train DNNs robustly with noisy labels. These noisy labels cause the performance degradation of DNNs due to the memorization effect by over-fitting. Earlier state-of-the-art methods used small loss tricks to efficiently resolve the robust training problem with noisy labels. In this paper, relationship between the uncertainties and the clean labels is analyzed. We present novel training method to use not only small loss trick but also labels that are likely to be clean labels selected from uncertainty called "Uncertain Aware Co-Training (UACT)". Our robust learning techniques (UACT) avoid over-fitting the DNNs by extremely noisy labels. By making better use of the uncertainty acquired from the network itself, we achieve good generalization performance. We compare the proposed method to the current state-of-the-art algorithms for noisy versions of MNIST, CIFAR-10, CIFAR-100, T-ImageNet and News to demonstrate its excellence.

Iterative confidence relabeling with deep ConvNets for organ segmentation with partial labels

Training deep ConvNets requires large labeled datasets. However, collecting pixel-level labels for medical image segmentation is very expensive and requires a high level of expertise. In addition, most existing segmentation masks provided by clinical experts focus on specific anatomical structures. In this paper, we propose a method dedicated to handle such partially labeled medical image datasets. We propose a strategy to identify pixels for which labels are correct, and to train Fully Convolutional Neural Networks with a multi-label loss adapted to this context. In addition, we introduce an iterative confidence self-training approach inspired by curriculum learning to relabel missing pixel labels, which relies on selecting the most confident prediction with a specifically designed confidence network that learns an uncertainty measure which is leveraged in our relabeling process. Our approach, INERRANT for Iterative coNfidencE Relabeling of paRtial ANnoTations, is thoroughly evaluated on two public datasets (TCAI and LITS), and one internal dataset with seven abdominal organ classes. We show that INERRANT robustly deals with partial labels, performing similarly to a model trained on all labels even for large missing label proportions. We also highlight the importance of our iterative learning scheme and the proposed confidence measure for optimal performance. Finally we show a practical use case where a limited number of completely labeled data are enriched by publicly available but partially labeled data.

CrowdTeacher: Robust Co-teaching with Noisy Answers and Sample-Specific Perturbations for Tabular Data

Samples with ground truth labels may not always be available in numerous domains. While learning from crowdsourcing labels has been explored, existing models can still fail in the presence of sparse, unreliable, or differing annotations. Co-teaching methods have shown promising improvements for computer vision problems with noisy labels by employing two classifiers trained on each others' confident samples in each batch. Inspired by the idea of separating confident and uncertain samples during the training process, we extend it for the crowdsourcing problem. Our model, CrowdTeacher, uses the idea that perturbation in the input space model can improve the robustness of the classifier for noisy labels. Treating crowdsourcing annotations as a source of noisy labeling, we perturb samples based on the certainty from the aggregated annotations. The perturbed samples are fed to a Co-teaching algorithm tuned to also accommodate smaller tabular data. We showcase the boost in predictive power attained using CrowdTeacher for both synthetic and real datasets across various label density settings. Our experiments reveal that our proposed approach beats baselines modeling individual annotations and then combining them, methods simultaneously learning a classifier and inferring truth labels, and the Co-teaching algorithm with aggregated labels through common truth inference methods.

Graph temporal ensembling based semi-supervised convolutional neural network with noisy labels for histopathology image analysis

Although convolutional neural networks have achieved tremendous success on histopathology image classification, they usually require large-scale clean annotated data and are sensitive to noisy labels. Unfortunately, labeling large-scale images is laborious, expensive and lowly reliable for pathologists. To address these problems, in this paper, we propose a novel self-ensembling based deep architecture to leverage the semantic information of annotated images and explore the information hidden in unlabeled data, and meanwhile being robust to noisy labels. Specifically, the proposed architecture first creates ensemble targets for feature and label predictions of training samples, by using exponential moving average (EMA) to aggregate feature and label predictions within multiple previous training epochs. Then, the ensemble targets within the same class are mapped into a cluster so that they are further enhanced. Next, a consistency cost is utilized to form consensus predictions under different configurations. Finally, we validate the proposed method with extensive experiments on lung and breast cancer datasets that contain thousands of images. It can achieve 90.5% and 89.5% image classification accuracy using only 20% labeled patients on the two datasets, respectively. This performance is comparable to that of the baseline method with all labeled patients. Experiments also demonstrate its robustness to small percentage of noisy labels.