Semi-Supervised Medical Image Classification With Relation-Driven Self-Ensembling Model

Multi-layer pseudo-supervision for histopathology tissue semantic segmentation using patch-level classification labels

Tissue-level semantic segmentation is a vital step in computational pathology. Fully-supervised models have already achieved outstanding performance with dense pixel-level annotations. However, drawing such labels on the giga-pixel whole slide images is extremely expensive and time-consuming. In this paper, we use only patch-level classification labels to achieve tissue semantic segmentation on histopathology images, finally reducing the annotation efforts. We propose a two-step model including a classification and a segmentation phases. In the classification phase, we propose a CAM-based model to generate pseudo masks by patch-level labels. In the segmentation phase, we achieve tissue semantic segmentation by our propose Multi-Layer Pseudo-Supervision. Several technical novelties have been proposed to reduce the information gap between pixel-level and patch-level annotations. As a part of this paper, we introduce a new weakly-supervised semantic segmentation (WSSS) dataset for lung adenocarcinoma (LUAD-HistoSeg). We conduct several experiments to evaluate our proposed model on two datasets. Our proposed model outperforms five state-of-the-art WSSS approaches. Note that we can achieve comparable quantitative and qualitative results with the fully-supervised model, with only around a 2% gap for MIoU and FwIoU. By comparing with manual labeling on a randomly sampled 100 patches dataset, patch-level labeling can greatly reduce the annotation time from hours to minutes. The source code and the released datasets are available at: https://github.com/ChuHan89/WSSS-Tissue.

Skin Lesion Analysis for Melanoma Detection Using the Novel Deep Learning Model Fuzzy GC-SCNN

Melanoma is easily detectable by visual examination since it occurs on the skin’s surface. In melanomas, which are the most severe types of skin cancer, the cells that make melanin are affected. However, the lack of expert opinion increases the processing time and cost of computer-aided skin cancer detection. As such, we aimed to incorporate deep learning algorithms to conduct automatic melanoma detection from dermoscopic images. The fuzzy-based GrabCut-stacked convolutional neural networks (GC-SCNN) model was applied for image training. The image features extraction and lesion classification were performed on different publicly available datasets. The fuzzy GC-SCNN coupled with the support vector machines (SVM) produced 99.75% classification accuracy and 100% sensitivity and specificity, respectively. Additionally, model performance was compared with existing techniques and outcomes suggesting the proposed model could detect and classify the lesion segments with higher accuracy and lower processing time than other techniques.

Few-Shot Learning with Collateral Location Coding and Single-Key Global Spatial Attention for Medical Image Classification

Humans are born with the ability to learn quickly by discerning objects from a few samples, to acquire new skills in a short period of time, and to make decisions based on limited prior experience and knowledge. The existing deep learning models for medical image classification often rely on a large number of labeled training samples, whereas the fast learning ability of deep neural networks has failed to develop. In addition, it requires a large amount of time and computing resource to retrain the model when the deep model encounters classes it has never seen before. However, for healthcare applications, enabling a model to generalize new clinical scenarios is of great importance. The existing image classification methods cannot explicitly use the location information of the pixel, making them insensitive to cues related only to the location. Besides, they also rely on local convolution and cannot properly utilize global information, which is essential for image classification. To alleviate these problems, we propose a collateral location coding to help the network explicitly exploit the location information of each pixel to make it easier for the network to recognize cues related to location only, and a single-key global spatial attention is designed to make the pixels at each location perceive the global spatial information in a low-cost way. Experimental results on three medical image benchmark datasets demonstrate that our proposed algorithm outperforms the state-of-the-art approaches in both effectiveness and generalization ability.

InClass nets: independent classifier networks for nonparametric estimation of conditional independence mixture models and unsupervised classification

Conditional independence mixture models (CIMMs) are an important class of statistical models used in many fields of science. We introduce a novel unsupervised machine learning technique called the independent classifier networks (InClass nets) technique for the nonparameteric estimation of CIMMs. InClass nets consist of multiple independent classifier neural networks (NNs), which are trained simultaneously using suitable cost functions. Leveraging the ability of NNs to handle high-dimensional data, the conditionally independent variates of the model are allowed to be individually high-dimensional, which is the main advantage of the proposed technique over existing non-machine-learning-based approaches. Two new theorems on the nonparametric identifiability of bivariate CIMMs are derived in the form of a necessary and a (different) sufficient condition for a bivariate CIMM to be identifiable. We use the InClass nets technique to perform CIMM estimation successfully for several examples. We provide a public implementation as a Python package called RainDancesVI.

A multimodal transformer to fuse images and metadata for skin disease classification

Skin disease cases are rising in prevalence, and the diagnosis of skin diseases is always a challenging task in the clinic. Utilizing deep learning to diagnose skin diseases could help to meet these challenges. In this study, a novel neural network is proposed for the classification of skin diseases. Since the datasets for the research consist of skin disease images and clinical metadata, we propose a novel multimodal Transformer, which consists of two encoders for both images and metadata and one decoder to fuse the multimodal information. In the proposed network, a suitable Vision Transformer (ViT) model is utilized as the backbone to extract image deep features. As for metadata, they are regarded as labels and a new Soft Label Encoder (SLE) is designed to embed them. Furthermore, in the decoder part, a novel Mutual Attention (MA) block is proposed to better fuse image features and metadata features. To evaluate the model's effectiveness, extensive experiments have been conducted on the private skin disease dataset and the benchmark dataset ISIC 2018. Compared with state-of-the-art methods, the proposed model shows better performance and represents an advancement in skin disease diagnosis.

Augmentation-Consistent Clustering Network for Diabetic Retinopathy Grading with Fewer Annotations

Diabetic retinopathy (DR) is currently one of the severe complications leading to blindness, and computer-aided, diagnosis technology-assisted DR grading has become a popular research trend especially for the development of deep learning methods. However, most deep learning-based DR grading models require a large number of annotations to provide data guidance, and it is laborious for experts to find subtle lesion areas from fundus images, making accurate annotation more expensive than other vision tasks. In contrast, large-scale unlabeled data are easily accessible, becoming a potential solution to reduce the annotating workload in DR grading. Thus, this paper explores the internal correlations from unknown fundus images assisted by limited labeled fundus images to solve the semisupervised DR grading problem and proposes an augmentation-consistent clustering network (ACCN) to address the above-mentioned challenges. Specifically, the augmentation provides an efficient cue for the similarity information of unlabeled fundus images, assisting the supervision from the labeled data. By mining the consistent correlations from augmentation and raw images, the ACCN can discover subtle lesion features by clustering with fewer annotations. Experiments on Messidor and APTOS 2019 datasets show that the ACCN surpasses many state-of-the-art methods in a semisupervised manner.

Handling Imbalanced Data: Uncertainty-guided Virtual Adversarial Training with Batch Nuclear-norm Optimization for Semi-supervised Medical Image Classification

In many clinical settings, a lot of medical image datasets suffer from imbalance problems, which makes predictions of trained models to be biased toward majority classes. Semi-supervised Learning (SSL) algorithms trained with such imbalanced datasets become more problematic since pseudo-supervision of unlabeled data are generated from the model's biased predictions. To address these issues, in this work, we propose a novel semi-supervised deep learning method, i.e., uncertainty-guided virtual adversarial training (VAT) with batch nuclear-norm (BNN) optimization, for large-scale medical image classification. To effectively exploit useful information from both labeled and unlabeled data, we leverage VAT and BNN optimization to harness the underlying knowledge, which helps to improve discriminability, diversity and generalization of the trained models. More concretely, our network is trained by minimizing a combination of four types of losses, including a supervised cross-entropy loss, a BNN loss defined on the output matrix of labeled data batch (lBNN loss), a negative BNN loss defined on the output matrix of unlabeled data batch (uBNN loss), and a VAT loss on both labeled and unlabeled data. We additionally propose to use uncertainty estimation to filter out unlabeled samples near the decision boundary when computing the VAT loss. We conduct comprehensive experiments to evaluate the performance of our method on two publicly available datasets and one in-house collected dataset. The experimental results demonstrated that our method achieved better results than state-of-the-art SSL methods.

Dual resolution deep learning network with self-attention mechanism for classification and localisation of colorectal cancer in histopathological images

AIMS

Microscopic examination is a basic diagnostic technology for colorectal cancer (CRC), but it is very laborious. We developed a dual resolution deep learning network with self-attention mechanism (DRSANet) which combines context and details for CRC binary classification and localisation in whole slide images (WSIs), and as a computer-aided diagnosis (CAD) to improve the sensitivity and specificity of doctors' diagnosis.

METHODS

Representative regions of interest (ROI) of each tissue type were manually delineated in WSIs by pathologists. Based on the same coordinates of centre position, patches were extracted at different magnification levels from the ROI. Specifically, patches from low magnification level contain contextual information, while from high magnification level provide important details. A dual-inputs network was designed to learn context and details simultaneously, and self-attention mechanism was used to selectively learn different positions in the images to enhance the performance.

RESULTS

In classification task, DRSANet outperformed the benchmark networks which only depended on the high magnification patches on two test set. Furthermore, in localisation task, DRSANet demonstrated a better localisation capability of tumour area in WSI with less areas of misidentification.

CONCLUSIONS

We compared DRSANet with benchmark networks which only use the patches from high magnification level. Experimental results reveal that the performance of DRSANet is better than the benchmark networks. Both context and details should be considered in deep learning method.

A holistic overview of deep learning approach in medical imaging

Medical images are a rich source of invaluable necessary information used by clinicians. Recent technologies have introduced many advancements for exploiting the most of this information and use it to generate better analysis. Deep learning (DL) techniques have been empowered in medical images analysis using computer-assisted imaging contexts and presenting a lot of solutions and improvements while analyzing these images by radiologists and other specialists. In this paper, we present a survey of DL techniques used for variety of tasks along with the different medical image's modalities to provide critical review of the recent developments in this direction. We have organized our paper to provide significant contribution of deep leaning traits and learn its concepts, which is in turn helpful for non-expert in medical society. Then, we present several applications of deep learning (e.g., segmentation, classification, detection, etc.) which are commonly used for clinical purposes for different anatomical site, and we also present the main key terms for DL attributes like basic architecture, data augmentation, transfer learning, and feature selection methods. Medical images as inputs to deep learning architectures will be the mainstream in the coming years, and novel DL techniques are predicted to be the core of medical images analysis. We conclude our paper by addressing some research challenges and the suggested solutions for them found in literature, and also future promises and directions for further developments.

Intra- and Inter-Pair Consistency for Semi-Supervised Gland Segmentation

Accurate gland segmentation in histology tissue images is a critical but challenging task. Although deep models have demonstrated superior performance in medical image segmentation, they commonly require a large amount of annotated data, which are hard to obtain due to the extensive labor costs and expertise required. In this paper, we propose an intra- and inter-pair consistency-based semi-supervised (I²CS) model that can be trained on both labeled and unlabeled histology images for gland segmentation. Considering that each image contains glands and hence different images could potentially share consistent semantics in the feature space, we introduce a novel intra- and inter-pair consistency module to explore such consistency for learning with unlabeled data. It first characterizes the pixel-level relation between a pair of images in the feature space to create an attention map that highlights the regions with the same semantics but on different images. Then, it imposes a consistency constraint on the attention maps obtained from multiple image pairs, and thus filters low-confidence attention regions to generate refined attention maps that are then merged with original features to improve their representation ability. In addition, we also design an object-level loss to address the issues caused by touching glands. We evaluated our model against several recent gland segmentation methods and three typical semi-supervised methods on the GlaS and CRAG datasets. Our results not only demonstrate the effectiveness of the proposed due consistency module and Obj-Dice loss, but also indicate that the proposed I²CS model achieves state-of-the-art gland segmentation performance on both benchmarks.

Development of computer-aided model to differentiate COVID-19 from pulmonary edema in lung CT scan: EDECOVID-net

The efforts made to prevent the spread of COVID-19 face specific challenges in diagnosing COVID-19 patients and differentiating them from patients with pulmonary edema. Although systemically administered pulmonary vasodilators and acetazolamide are of great benefit for treating pulmonary edema, they should not be used to treat COVID-19 as they carry the risk of several adverse consequences, including worsening the matching of ventilation and perfusion, impaired carbon dioxide transport, systemic hypotension, and increased work of breathing. This study proposes a machine learning-based method (EDECOVID-net) that automatically differentiates the COVID-19 symptoms from pulmonary edema in lung CT scans using radiomic features. To the best of our knowledge, EDECOVID-net is the first method to differentiate COVID-19 from pulmonary edema and a helpful tool for diagnosing COVID-19 at early stages. The EDECOVID-net has been proposed as a new machine learning-based method with some advantages, such as having simple structure and few mathematical calculations. In total, 13 717 imaging patches, including 5759 COVID-19 and 7958 edema images, were extracted using a CT incision by a specialist radiologist. The EDECOVID-net can distinguish the patients with COVID-19 from those with pulmonary edema with an accuracy of 0.98. In addition, the accuracy of the EDECOVID-net algorithm is compared with other machine learning methods, such as VGG-16 (Acc = 0.94), VGG-19 (Acc = 0.96), Xception (Acc = 0.95), ResNet101 (Acc = 0.97), and DenseNet20l (Acc = 0.97).

Co-optimization Learning Network for MRI Segmentation of Ischemic Penumbra Tissues

Convolutional neural networks (CNNs) have brought hope for the medical image auxiliary diagnosis. However, the shortfall of labeled medical image data is the bottleneck that limits the performance improvement of supervised CNN methods. In addition, annotating a large number of labeled medical image data is often expensive and time-consuming. In this study, we propose a co-optimization learning network (COL-Net) for Magnetic Resonance Imaging (MRI) segmentation of ischemic penumbra tissues. COL-Net base on the limited labeled samples and consists of an unsupervised reconstruction network (R), a supervised segmentation network (S), and a transfer block (T). The reconstruction network extracts the robust features from reconstructing pseudo unlabeled samples, which is the auxiliary branch of the segmentation network. The segmentation network is used to segment the target lesions under the limited labeled samples and the auxiliary of the reconstruction network. The transfer block is used to co-optimization the feature maps between the bottlenecks of the reconstruction network and segmentation network. We propose a mix loss function to optimize COL-Net. COL-Net is verified on the public ischemic penumbra segmentation challenge (SPES) with two dozen labeled samples. Results demonstrate that COL-Net has high predictive accuracy and generalization with the Dice coefficient of 0.79. The extended experiment also shows COL-Net outperforms most supervised segmentation methods. COL-Net is a meaningful attempt to alleviate the limited labeled sample problem in medical image segmentation.

Improving the Classification Performance of Esophageal Disease on Small Dataset by Semi-supervised Efficient Contrastive Learning

The classification of esophageal disease based on gastroscopic images is important in the clinical treatment, and is also helpful in providing patients with follow-up treatment plans and preventing lesion deterioration. In recent years, deep learning has achieved many satisfactory results in gastroscopic image classification tasks. However, most of them need a training set that consists of large numbers of images labeled by experienced experts. To reduce the image annotation burdens and improve the classification ability on small labeled gastroscopic image datasets, this study proposed a novel semi-supervised efficient contrastive learning (SSECL) classification method for esophageal disease. First, an efficient contrastive pair generation (ECPG) module was proposed to generate efficient contrastive pairs (ECPs), which took advantage of the high similarity features of images from the same lesion. Then, an unsupervised visual feature representation containing the general feature of esophageal gastroscopic images is learned by unsupervised efficient contrastive learning (UECL). At last, the feature representation will be transferred to the down-stream esophageal disease classification task. The experimental results have demonstrated that the classification accuracy of SSECL is 92.57%, which is better than that of the other state-of-the-art semi-supervised methods and is also higher than the classification method based on transfer learning (TL) by 2.28%. Thus, SSECL has solved the challenging problem of improving the classification result on small gastroscopic image dataset by fully utilizing the unlabeled gastroscopic images and the high similarity information among images from the same lesion. It also brings new insights into medical image classification tasks.

Accurate recognition of colorectal cancer with semi-supervised deep learning on pathological images

Machine-assisted pathological recognition has been focused on supervised learning (SL) that suffers from a significant annotation bottleneck. We propose a semi-supervised learning (SSL) method based on the mean teacher architecture using 13,111 whole slide images of colorectal cancer from 8803 subjects from 13 independent centers. SSL (~3150 labeled, ~40,950 unlabeled; ~6300 labeled, ~37,800 unlabeled patches) performs significantly better than the SL. No significant difference is found between SSL (~6300 labeled, ~37,800 unlabeled) and SL (~44,100 labeled) at patch-level diagnoses (area under the curve (AUC): 0.980 ± 0.014 vs. 0.987 ± 0.008, P value = 0.134) and patient-level diagnoses (AUC: 0.974 ± 0.013 vs. 0.980 ± 0.010, P value = 0.117), which is close to human pathologists (average AUC: 0.969). The evaluation on 15,000 lung and 294,912 lymph node images also confirm SSL can achieve similar performance as that of SL with massive annotations. SSL dramatically reduces the annotations, which has great potential to effectively build expert-level pathological artificial intelligence platforms in practice.

RCTE: A reliable and consistent temporal-ensembling framework for semi-supervised segmentation of COVID-19 lesions

The segmentation of COVID-19 lesions from computed tomography (CT) scans is crucial to develop an efficient automated diagnosis system. Deep learning (DL) has shown success in different segmentation tasks. However, an efficient DL approach requires a large amount of accurately annotated data, which is difficult to aggregate owing to the urgent situation of COVID-19. Inaccurate annotation can easily occur without experts, and segmentation performance is substantially worsened by noisy annotations. Therefore, this study presents a reliable and consistent temporal-ensembling (RCTE) framework for semi-supervised lesion segmentation. A segmentation network is integrated into a teacher-student architecture to segment infection regions from a limited number of annotated CT scans and a large number of unannotated CT scans. The network generates reliable and unreliable targets, and to evenly handle these targets potentially degrades performance. To address this, a reliable teacher-student architecture is introduced, where a reliable teacher network is the exponential moving average (EMA) of a reliable student network that is reliably renovated by restraining the student involvement to EMA when its loss is larger. We also present a noise-aware loss based on improvements to generalized cross-entropy loss to lead the segmentation performance toward noisy annotations. Comprehensive analysis validates the robustness of RCTE over recent cutting-edge semi-supervised segmentation techniques, with a 65.87% Dice score.

Exploiting Shared Knowledge From Non-COVID Lesions for Annotation-Efficient COVID-19 CT Lung Infection Segmentation

The novel Coronavirus disease (COVID-19) is a highly contagious virus and has spread all over the world, posing an extremely serious threat to all countries. Automatic lung infection segmentation from computed tomography (CT) plays an important role in the quantitative analysis of COVID-19. However, the major challenge lies in the inadequacy of annotated COVID-19 datasets. Currently, there are several public non-COVID lung lesion segmentation datasets, providing the potential for generalizing useful information to the related COVID-19 segmentation task. In this paper, we propose a novel relation-driven collaborative learning model to exploit shared knowledge from non-COVID lesions for annotation-efficient COVID-19 CT lung infection segmentation. The model consists of a general encoder to capture general lung lesion features based on multiple non-COVID lesions, and a target encoder to focus on task-specific features based on COVID-19 infections. We develop a collaborative learning scheme to regularize feature-level relation consistency of given input and encourage the model to learn more general and discriminative representation of COVID-19 infections. Extensive experiments demonstrate that trained with limited COVID-19 data, exploiting shared knowledge from non-COVID lesions can further improve state-of-the-art performance with up to 3.0% in dice similarity coefficient and 4.2% in normalized surface dice. In addition, experimental results on large scale 2D dataset with CT slices show that our method significantly outperforms cutting-edge segmentation methods metrics. Our method promotes new insights into annotation-efficient deep learning and illustrates strong potential for real-world applications in the global fight against COVID-19 in the absence of sufficient high-quality annotations.

Multi-modal Broad Learning System for Medical Image and Text-based Classification

Automatic classification of medical images plays an essential role in computer-aided diagnosis. However, the medical images arise from the small number of available data and the improvement of existing data-enhancement methods are limited. In order to fulfil this demand, a Multi-Modal Broad Learning System (M²-BLS) is proposed, which has two subnetworks for simultaneous learning of both medical images and the corresponding radiology reports. M²-BLS provides two advantages: i) our M²-BLS has closed-form solution and avoids iterative training, once the image feature is available; ii) benefit from the simultaneous learning of both image and text data, our M²-BLS achieves high accuracy for medical classification. Experimental results on the publicly available datasets IU X-RAY and PEIR GROSS_895 show that our M²-BLS highly improves the classification performance, compared to SOTA deep models that learn single-type of data information only.

Self-supervised driven consistency training for annotation efficient histopathology image analysis

Training a neural network with a large labeled dataset is still a dominant paradigm in computational histopathology. However, obtaining such exhaustive manual annotations is often expensive, laborious, and prone to inter and intra-observer variability. While recent self-supervised and semi-supervised methods can alleviate this need by learning unsupervised feature representations, they still struggle to generalize well to downstream tasks when the number of labeled instances is small. In this work, we overcome this challenge by leveraging both task-agnostic and task-specific unlabeled data based on two novel strategies: (i) a self-supervised pretext task that harnesses the underlying multi-resolution contextual cues in histology whole-slide images to learn a powerful supervisory signal for unsupervised representation learning; (ii) a new teacher-student semi-supervised consistency paradigm that learns to effectively transfer the pretrained representations to downstream tasks based on prediction consistency with the task-specific unlabeled data. We carry out extensive validation experiments on three histopathology benchmark datasets across two classification and one regression based tasks, i.e., tumor metastasis detection, tissue type classification, and tumor cellularity quantification. Under limited-label data, the proposed method yields tangible improvements, which is close to or even outperforming other state-of-the-art self-supervised and supervised baselines. Furthermore, we empirically show that the idea of bootstrapping the self-supervised pretrained features is an effective way to improve the task-specific semi-supervised learning on standard benchmarks. Code and pretrained models are made available at: https://github.com/srinidhiPY/SSL_CR_Histo.

DR-Net: dual-rotation network with feature map enhancement for medical image segmentation

To obtain more semantic information with small samples for medical image segmentation, this paper proposes a simple and efficient dual-rotation network (DR-Net) that strengthens the quality of both local and global feature maps. The key steps of the DR-Net algorithm are as follows (as shown in Fig. 1). First, the number of channels in each layer is divided into four equal portions. Then, different rotation strategies are used to obtain a rotation feature map in multiple directions for each subimage. Then, the multiscale volume product and dilated convolution are used to learn the local and global features of feature maps. Finally, the residual strategy and integration strategy are used to fuse the generated feature maps. Experimental results demonstrate that the DR-Net method can obtain higher segmentation accuracy on both the CHAOS and BraTS data sets compared to the state-of-the-art methods.

Learning From Multiple Datasets With Heterogeneous and Partial Labels for Universal Lesion Detection in CT

Large-scale datasets with high-quality labels are desired for training accurate deep learning models. However, due to the annotation cost, datasets in medical imaging are often either partially-labeled or small. For example, DeepLesion is such a large-scale CT image dataset with lesions of various types, but it also has many unlabeled lesions (missing annotations). When training a lesion detector on a partially-labeled dataset, the missing annotations will generate incorrect negative signals and degrade the performance. Besides DeepLesion, there are several small single-type datasets, such as LUNA for lung nodules and LiTS for liver tumors. These datasets have heterogeneous label scopes, i.e., different lesion types are labeled in different datasets with other types ignored. In this work, we aim to develop a universal lesion detection algorithm to detect a variety of lesions. The problem of heterogeneous and partial labels is tackled. First, we build a simple yet effective lesion detection framework named Lesion ENSemble (LENS). LENS can efficiently learn from multiple heterogeneous lesion datasets in a multi-task fashion and leverage their synergy by proposal fusion. Next, we propose strategies to mine missing annotations from partially-labeled datasets by exploiting clinical prior knowledge and cross-dataset knowledge transfer. Finally, we train our framework on four public lesion datasets and evaluate it on 800 manually-labeled sub-volumes in DeepLesion. Our method brings a relative improvement of 49% compared to the current state-of-the-art approach in the metric of average sensitivity. We have publicly released our manual 3D annotations of DeepLesion online.^{1 1}https://github.com/viggin/DeepLesion_manual_test_set.

Learning Inductive Attention Guidance for Partially Supervised Pancreatic Ductal Adenocarcinoma Prediction

Pancreatic ductal adenocarcinoma (PDAC) is the third most common cause of cancer death in the United States. Predicting tumors like PDACs (including both classification and segmentation) from medical images by deep learning is becoming a growing trend, but usually a large number of annotated data are required for training, which is very labor-intensive and time-consuming. In this paper, we consider a partially supervised setting, where cheap image-level annotations are provided for all the training data, and the costly per-voxel annotations are only available for a subset of them. We propose an Inductive Attention Guidance Network (IAG-Net) to jointly learn a global image-level classifier for normal/PDAC classification and a local voxel-level classifier for semi-supervised PDAC segmentation. We instantiate both the global and the local classifiers by multiple instance learning (MIL), where the attention guidance, indicating roughly where the PDAC regions are, is the key to bridging them: For global MIL based normal/PDAC classification, attention serves as a weight for each instance (voxel) during MIL pooling, which eliminates the distraction from the background; For local MIL based semi-supervised PDAC segmentation, the attention guidance is inductive, which not only provides bag-level pseudo-labels to training data without per-voxel annotations for MIL training, but also acts as a proxy of an instance-level classifier. Experimental results show that our IAG-Net boosts PDAC segmentation accuracy by more than 5% compared with the state-of-the-arts.

Student becomes teacher: training faster deep learning lightweight networks for automated identification of optical coherence tomography B-scans of interest using a student-teacher framework

This work explores a student-teacher framework that leverages unlabeled images to train lightweight deep learning models with fewer parameters to perform fast automated detection of optical coherence tomography B-scans of interest. Twenty-seven lightweight models (LWMs) from four families of models were trained on expert-labeled B-scans (∼70 K) as either "abnormal" or "normal", which established a baseline performance for the models. Then the LWMs were trained from random initialization using a student-teacher framework to incorporate a large number of unlabeled B-scans (∼500 K). A pre-trained ResNet50 model served as the teacher network. The ResNet50 teacher model achieved 96.0% validation accuracy and the validation accuracy achieved by the LWMs ranged from 89.6% to 95.1%. The best performing LWMs were 2.53 to 4.13 times faster than ResNet50 (0.109s to 0.178s vs. 0.452s). All LWMs benefitted from increasing the training set by including unlabeled B-scans in the student-teacher framework, with several models achieving validation accuracy of 96.0% or higher. The three best-performing models achieved comparable sensitivity and specificity in two hold-out test sets to the teacher network. We demonstrated the effectiveness of a student-teacher framework for training fast LWMs for automated B-scan of interest detection leveraging unlabeled, routinely-available data.

A primer on deep learning and convolutional neural networks for clinicians

Deep learning is nowadays at the forefront of artificial intelligence. More precisely, the use of convolutional neural networks has drastically improved the learning capabilities of computer vision applications, being able to directly consider raw data without any prior feature extraction. Advanced methods in the machine learning field, such as adaptive momentum algorithms or dropout regularization, have dramatically improved the convolutional neural networks predicting ability, outperforming that of conventional fully connected neural networks. This work summarizes, in an intended didactic way, the main aspects of these cutting-edge techniques from a medical imaging perspective.

Medical Instrument Segmentation in 3D US by Hybrid Constrained Semi-Supervised Learning

Medical instrument segmentation in 3D ultrasound is essential for image-guided intervention. However, to train a successful deep neural network for instrument segmentation, a large number of labeled images are required, which is expensive and time-consuming to obtain. In this article, we propose a semi-supervised learning (SSL) framework for instrument segmentation in 3D US, which requires much less annotation effort than the existing methods. To achieve the SSL learning, a Dual-UNet is proposed to segment the instrument. The Dual-UNet leverages unlabeled data using a novel hybrid loss function, consisting of uncertainty and contextual constraints. Specifically, the uncertainty constraints leverage the uncertainty estimation of the predictions of the UNet, and therefore improve the unlabeled information for SSL training. In addition, contextual constraints exploit the contextual information of the training images, which are used as the complementary information for voxel-wise uncertainty estimation. Extensive experiments on multiple ex-vivo and in-vivo datasets show that our proposed method achieves Dice score of about 68.6%-69.1% and the inference time of about 1 sec. per volume. These results are better than the state-of-the-art SSL methods and the inference time is comparable to the supervised approaches.

Deep Learning Application for Analyzing of Constituents and Their Correlations in the Interpretations of Medical Images

The need for time and attention, given by the doctor to the patient, due to the increased volume of medical data to be interpreted and filtered for diagnostic and therapeutic purposes has encouraged the development of the option to support, constructively and effectively, deep learning models. Deep learning (DL) has experienced an exponential development in recent years, with a major impact on interpretations of the medical image. This has influenced the development, diversification and increase of the quality of scientific data, the development of knowledge construction methods and the improvement of DL models used in medical applications. All research papers focus on description, highlighting, classification of one of the constituent elements of deep learning models (DL), used in the interpretation of medical images and do not provide a unified picture of the importance and impact of each constituent in the performance of DL models. The novelty in our paper consists primarily in the unitary approach, of the constituent elements of DL models, namely, data, tools used by DL architectures or specifically constructed DL architecture combinations and highlighting their "key" features, for completion of tasks in current applications in the interpretation of medical images. The use of "key" characteristics specific to each constituent of DL models and the correct determination of their correlations, may be the subject of future research, with the aim of increasing the performance of DL models in the interpretation of medical images.

Semi-supervised classification of radiology images with NoTeacher: A teacher that is not mean

Deep learning models achieve strong performance for radiology image classification, but their practical application is bottle-necked by the need for large labeled training datasets. Semi-supervised learning (SSL) approaches leverage small labeled datasets alongside larger unlabeled datasets and offer potential for reducing labeling cost. In this work, we introduce NoTeacher, a novel consistency-based SSL framework which incorporates probabilistic graphical models. Unlike Mean Teacher which maintains a teacher network updated via a temporal ensemble, NoTeacher employs two independent networks, thereby eliminating the need for a teacher network. We demonstrate how NoTeacher can be customized to handle a range of challenges in radiology image classification. Specifically, we describe adaptations for scenarios with 2D and 3D inputs, uni and multi-label classification, and class distribution mismatch between labeled and unlabeled portions of the training data. In realistic empirical evaluations on three public benchmark datasets spanning the workhorse modalities of radiology (X-Ray, CT, MRI), we show that NoTeacher achieves over 90-95% of the fully supervised AUROC with less than 5-15% labeling budget. Further, NoTeacher outperforms established SSL methods with minimal hyperparameter tuning, and has implications as a principled and practical option for semi-supervised learning in radiology applications.

MS-ANet: deep learning for automated multi-label thoracic disease detection and classification

The chest X-ray is one of the most common radiological examination types for the diagnosis of chest diseases. Nowadays, the automatic classification technology of radiological images has been widely used in clinical diagnosis and treatment plans. However, each disease has its own different response characteristic receptive field region, which is the main challenge for chest disease classification tasks. Besides, the imbalance of sample data categories further increases the difficulty of tasks. To solve these problems, we propose a new multi-label chest disease image classification scheme based on a multi-scale attention network. In this scheme, multi-scale information is iteratively fused to focus on regions with a high probability of disease, to effectively mine more meaningful information from data. A novel loss function is also designed to improve the rationality of visual perception and multi-label image classification, which forces the consistency of attention regions before and after image transformation. A comprehensive experiment was carried out on the Chest X-Ray14 and CheXpert datasets, separately containing over 100,000 frontal-view and 200,000 front and side view X-ray images with 14 diseases. The AUROC is 0.850 and 0.815 respectively on the two data sets, which achieve the state-of-the-art results, verified the effectiveness of this method in chest X-ray image classification. This study has important practical significance for using AI algorithms to assist radiologists in improving work efficiency and diagnostic accuracy.

Federated semi-supervised learning for COVID region segmentation in chest CT using multi-national data from China, Italy, Japan

The recent outbreak of Coronavirus Disease 2019 (COVID-19) has led to urgent needs for reliable diagnosis and management of SARS-CoV-2 infection. The current guideline is using RT-PCR for testing. As a complimentary tool with diagnostic imaging, chest Computed Tomography (CT) has been shown to be able to reveal visual patterns characteristic for COVID-19, which has definite value at several stages during the disease course. To facilitate CT analysis, recent efforts have focused on computer-aided characterization and diagnosis with chest CT scan, which has shown promising results. However, domain shift of data across clinical data centers poses a serious challenge when deploying learning-based models. A common way to alleviate this issue is to fine-tune the model locally with the target domains local data and annotations. Unfortunately, the availability and quality of local annotations usually varies due to heterogeneity in equipment and distribution of medical resources across the globe. This impact may be pronounced in the detection of COVID-19, since the relevant patterns vary in size, shape, and texture. In this work, we attempt to find a solution for this challenge via federated and semi-supervised learning. A multi-national database consisting of 1704 scans from three countries is adopted to study the performance gap, when training a model with one dataset and applying it to another. Expert radiologists manually delineated 945 scans for COVID-19 findings. In handling the variability in both the data and annotations, a novel federated semi-supervised learning technique is proposed to fully utilize all available data (with or without annotations). Federated learning avoids the need for sensitive data-sharing, which makes it favorable for institutions and nations with strict regulatory policy on data privacy. Moreover, semi-supervision potentially reduces the annotation burden under a distributed setting. The proposed framework is shown to be effective compared to fully supervised scenarios with conventional data sharing instead of model weight sharing.

A bird's-eye view of deep learning in bioimage analysis

Deep learning of artificial neural networks has become the de facto standard approach to solving data analysis problems in virtually all fields of science and engineering. Also in biology and medicine, deep learning technologies are fundamentally transforming how we acquire, process, analyze, and interpret data, with potentially far-reaching consequences for healthcare. In this mini-review, we take a bird's-eye view at the past, present, and future developments of deep learning, starting from science at large, to biomedical imaging, and bioimage analysis in particular.