Multiple Instance Learning Research Articles

Annotation protocol and crowdsourcing multiple instance learning classification of skin histological images: The CR-AI4SkIN dataset

Digital Pathology (DP) has experienced a significant growth in recent years and has become an essential tool for diagnosing and prognosis of tumors. The availability of Whole Slide Images (WSIs) and the implementation of Deep Learning (DL) algorithms have paved the way for the appearance of Artificial Intelligence (AI) systems that support the diagnosis process. These systems require extensive and varied data for their training to be successful. However, creating labeled datasets in histopathology is laborious and time-consuming. We have developed a crowdsourcing-multiple instance labeling/learning protocol that is applied to the creation and use of the CR-AI4SkIN dataset.22For further information on the code and dataset: https://github.com/wizmik12/Crowdsourcing-MIL-Skin-Cancer/. CR-AI4SkIN contains 271 WSIs of 7 Cutaneous Spindle Cell (CSC) neoplasms with expert and non-expert labels at region and WSI levels. It is the first dataset of these types of neoplasms made available. The regions selected by the experts are used to learn an automatic extractor of Regions of Interest (ROIs) from WSIs. To produce the embedding of each WSI, the representations of patches within the ROIs are obtained using a contrastive learning method, and then combined. Finally, they are fed to a Gaussian process-based crowdsourcing classifier, which utilizes the noisy non-expert WSI labels. We validate our crowdsourcing-multiple instance learning method in the CR-AI4SkIN dataset, addressing a binary classification problem (malign vs. benign). The proposed method obtains an F1 score of 0.7911 on the test set, outperforming three widely used aggregation methods for crowdsourcing tasks. Furthermore, our crowdsourcing method also outperforms the supervised model with expert labels on the test set (F1-score = 0.6035). The promising results support the proposed crowdsourcing multiple instance learning annotation protocol. It also validates the automatic extraction of interest regions and the use of contrastive embedding and Gaussian process classification to perform crowdsourcing classification tasks.

A hybrid few-shot multiple-instance learning model predicting the aggressiveness of lymphoma in PET/CT images

Background and objectivePatients with aggressive non-Hodgkin lymphoma (NHL) undergo distinct therapy strategies compared with indolent NHL patients. However, it is challenging to estimate NHL aggressiveness based on visual inspection of positron emission tomography (PET) or computed tomography (CT) images. Since diffuse large B-cell lymphoma (DLBCL) and Follicular lymphoma (FL) are the most typical and dominant aggressive and indolent NHL, respectively, this study aims to develop an artificial-intelligence-enabled model to distinguish DLBCL from FL in PET/CT images as the first step to tackle this challenge. MethodsWe propose a hybrid few-shot multiple-instance learning model to predict the aggressiveness of the NHL. First, rotation-based self-supervision learning (SSL) has been employed to train the encoder on a large-scale, publicly available CT image dataset. Second, hybrid instance-level features are obtained for each NHL lesion by combining deep features with the radiomics features from both PET and CT modalities. Third, instance-level features are transformed into bag-level (or patient-level) representations. Finally, bag-level representations are fed into a distance-based classifier through few-shot learning to predict NHL aggressiveness. ResultsOur model achieves an accuracy of 0.751 ± 0.008, a sensitivity of 0.787 ± 0.012, a specificity of 0.715 ± 0.013, an F1-score of 0.753 ± 0.009, and an area under the curve (AUC) of 0.795 ± 0.009 at the bag level. It outperforms the typical counterparts that use the radiomic features, random forest for feature selection, and support vector machines (SVMs) as classifiers. The three counterparts yield accuracies of 0.714 ± 0.023, 0.705 ± 0.008, and 0.698 ± 0.008, respectively. Moreover, settings of the SSL training dataset (Deep lesion) and task (rotation), hybrid CT and radiomic PET features, the pool-layer strategy of maximum, and distance-based classifier generate the best model. ConclusionsA hybrid few-shot multiple-instance learning model can predict lymphoma aggressiveness in PET/CT images and could be a potential tool for determining therapy strategies. Hybrid features and the combination of SSL, few-shot learning, and weakly supervised learning are the two powerful pillars of the model, and these can be expanded to other medical applications with limited samples and incomplete annotations.

Whole slide image representation in bone marrow cytology

One of the goals of AI-based computational pathology is to generate compact representations of whole slide images (WSIs) that capture the essential information needed for diagnosis. While such approaches have been applied to histopathology, few applications have been reported in cytology. Bone marrow aspirate cytology is the basis for key clinical decisions in hematology. However, visual inspection of aspirate specimens is a tedious and complex process subject to variation in interpretation, and hematopathology expertise is scarce. The ability to generate a compact representation of an aspirate specimen may form the basis for clinical decision-support tools in hematology. In this study, we leverage our previously published end-to-end AI-based system for counting and classifying cells from bone marrow aspirate WSIs, which enables the direct use of individual cells as inputs rather than WSI patches. We then construct bags of individual cell features from each WSI, and apply multiple instance learning to extract their vector representations. To evaluate the quality of our representations, we conducted WSI retrieval and classification tasks. Our results show that we achieved a mAP@10 of 0.58 ±0.02 in WSI-level image retrieval, surpassing the random-retrieval baseline of 0.39 ±0.1. Furthermore, we predicted five diagnostic labels for individual aspirate WSIs with a weighted-average F1 score of 0.57 ±0.03 using a k-nearest-neighbors (k-NN) model, outperforming guessing using empirical class prior probabilities (0.26 ±0.02). We present the first example of exploring trainable mechanisms to generate compact, slide-level representations in bone marrow cytology with deep learning. This method has the potential to summarize complex semantic information in WSIs toward improved diagnostics in hematology, and may eventually support AI-assisted computational pathology approaches.

Deep Learning Models Capture Histological Disease Activity in Crohn's Disease and Ulcerative Colitis with High Fidelity.

Histological disease activity in inflammatory bowel disease [IBD] is associated with clinical outcomes and is an important endpoint in drug development. We developed deep learning models for automating histological assessments in IBD. Histology images of intestinal mucosa from phase 2 and phase 3 clinical trials in Crohn's disease [CD] and ulcerative colitis [UC] were used to train artificial intelligence [AI] models to predict the Global Histology Activity Score [GHAS] for CD and Geboes histopathology score for UC. Three AI methods were compared. AI models were evaluated on held-back testing sets, and model predictions were compared against an expert central reader and five independent pathologists. The model based on multiple instance learning and the attention mechanism [SA-AbMILP] demonstrated the best performance among competing models. AI-modelled GHAS and Geboes subgrades matched central readings with moderate to substantial agreement, with accuracies ranging from 65% to 89%. Furthermore, the model was able to distinguish the presence and absence of pathology across four selected histological features, with accuracies for colon in both CD and UC ranging from 87% to 94% and for CD ileum ranging from 76% to 83%. For both CD and UC and across anatomical compartments [ileum and colon] in CD, comparable accuracies against central readings were found between the model-assigned scores and scores by an independent set of pathologists. Deep learning models based upon GHAS and Geboes scoring systems were effective at distinguishing between the presence and absence of IBD microscopic disease activity.

Local Self-attention-based Hybrid Multiple Instance Learning for Partial Spoof Speech Detection

The development of speech synthesis technology has increased the attention toward the threat of spoofed speech. Although various high-performance spoofing countermeasures have been proposed in recent years, a particular scenario is overlooked: partially spoofed audio, where spoofed utterances may contain both spoofed and bona fide segments. Currently, the research on partially spoofed speech detection is lacking. The existing methods either train with partially spoofed speech at utterance level, resulting in gradient conflicting at the segment level, or directly train with segment level data, which requires segment labels that are difficult to obtain in practice. In this study, to better detect partially spoofed speech when only utterance labels are available, we formulate partially spoofed speech detection into a multiple instance learning (MIL) problem. The typical MIL uses a pooling layer to fuse patch scores as a whole, and we propose a hybrid MIL (H-MIL) framework based on max and log-sum-exp pooling methods, which can learn better segment representations to improve partially spoofed speech detection performance. Theoretical and experimental verification shows that H-MIL can effectively relieve the gradient conflicting and gradient vanishing problems. In addition, we analyze the local correlations between segments and introduce a local self-attention mechanism to enhance segment features, which further promotes the detection performance. In our experiments, we provide not only detection results at the segment and utterance levels but also some detailed visualization analysis, including the effect of spoof ratio and cross-dataset detection. The experimental results demonstrate the effective detection performance of our method at both the utterance and segment levels, especially when dealing with low spoof ratio attacks. The results confirm that our approach can better deal with partially spoofed speech detection than previous methods.

An image caption model based on attention mechanism and deep reinforcement learning.

Image caption technology aims to convert visual features of images, extracted by computers, into meaningful semantic information. Therefore, the computers can generate text descriptions that resemble human perception, enabling tasks such as image classification, retrieval, and analysis. In recent years, the performance of image caption has been significantly enhanced with the introduction of encoder-decoder architecture in machine translation and the utilization of deep neural networks. However, several challenges still persist in this domain. Therefore, this paper proposes a novel method to address the issue of visual information loss and non-dynamic adjustment of input images during decoding. We introduce a guided decoding network that establishes a connection between the encoding and decoding parts. Through this connection, encoding information can provide guidance to the decoding process, facilitating automatic adjustment of the decoding information. In addition, Dense Convolutional Network (DenseNet) and Multiple Instance Learning (MIL) are adopted in the image encoder, and Nested Long Short-Term Memory (NLSTM) is utilized as the decoder to enhance the extraction and parsing capability of image information during the encoding and decoding process. In order to further improve the performance of our image caption model, this study incorporates an attention mechanism to focus details and constructs a double-layer decoding structure, which facilitates the enhancement of the model in terms of providing more detailed descriptions and enriched semantic information. Furthermore, the Deep Reinforcement Learning (DRL) method is employed to train the model by directly optimizing the identical set of evaluation indexes, which solves the problem of inconsistent training and evaluation standards. Finally, the model is trained and tested on MS COCO and Flickr 30 k datasets, and the results show that the model has improved compared with commonly used models in the evaluation indicators such as BLEU, METEOR and CIDEr.

Data-efficient machine learning methods in the ME-TIME study: Rationale and design of a longitudinal study to detect atrial fibrillation and heart failure from wearables

Smartwatches enable continuous and noninvasive time series monitoring of cardiovascular biomarkers like heart rate (from photoplethysmograms), step counter, skin temperature, et cetera; as such, they have promise in assisting in early detection and prevention of cardiovascular disease. Although these biomarkers may not be directly useful to physicians, a machine learning (ML) model could find clinically relevant patterns. Unfortunately, ML models typically need supervised (ie, annotated) data, and labeling of large amounts of continuous data is very labor intensive. Therefore, ML methods that are data efficient, ie, needing a low number of labels, are required to detect potential clinical value in patterns found in wearable data. The primary study objective of the ME-TIME (Machine Learning Enabled Time Series Analysis in Medicine) study is to design an ML model that can detect atrial fibrillation (AF) and heart failure (HF) from wearable data in a data-efficient manner. To achieve this, self-supervised and weakly supervised learning techniques are used. Two hundred subjects (100 reference, 50 AF, and 50 HF) are being invited to participate in wearing a Fitbit fitness tracker for 3 months. Interested volunteers are sent a questionnaire to determine their health, in particular cardiovascular health. Volunteers without any (history of) serious illness are assigned to the reference group. Participants with AF and HF are recruited in the Haga teaching hospital in The Hague, The Netherlands. Enrollment commenced on May 1, 2022, and as of the time of this report, 62 subjects have been included in the study. Preliminary analysis of the data reveals significant inter-subject variability. Notably, we identified heart rate recovery curves and time-delayed correlations between heart rate and step count as potential strong indicators for heart disease. Using self-supervised and multiple-instance learning techniques, we hypothesize that patterns specific to AF and HF can be found in continuous data obtained from smartwatches.

A Structure-aware Hierarchical Graph-based Multiple Instance Learning Framework for pT Staging in Histopathological Image.

Pathological primary tumor (pT) stage focuses on the infiltration degree of the primary tumor to surrounding tissues, which relates to the prognosis and treatment choices. The pT staging relies on the field-of-views from multiple magnifications in the gigapixel images, which makes pixel-level annotation difficult. Therefore, this task is usually formulated as a weakly supervised whole slide image (WSI) classification task with the slide-level label. Existing weakly-supervised classification methods mainly follow the multiple instance learning paradigm, which takes the patches from single magnification as the instances and extracts their morphological features independently. However, they cannot progressively represent the contextual information from multiple magnifications, which is critical for pT staging. Therefore, we propose a structure-aware hierarchical graph-based multi-instance learning framework (SGMF) inspired by the diagnostic process of pathologists. Specifically, a novel graph-based instance organization method is proposed, namely structure-aware hierarchical graph (SAHG), to represent the WSI. Based on that, we design a novel hierarchical attention-based graph representation (HAGR) network to capture the critical patterns for pT staging by learning cross-scale spatial features. Finally, the top nodes of SAHG are aggregated by a global attention layer for bag-level representation. Extensive studies on three large-scale multi-center pT staging datasets with two different cancer types demonstrate the effectiveness of SGMF, which outperforms state-of-the-art up to 5.6% in the F1 score.

Multi-scale representation attention based deep multiple instance learning for gigapixel whole slide image analysis.

Recently, convolutional neural networks (CNNs) directly using whole slide images (WSIs) for tumor diagnosis and analysis have attracted considerable attention, because they only utilize the slide-level label for model training without any additional annotations. However, it is still a challenging task to directly handle gigapixel WSIs, due to the billions of pixels and intra-variations in each WSI. To overcome this problem, in this paper, we propose a novel end-to-end interpretable deep MIL framework for WSI analysis, by using a two-branch deep neural network and a multi-scale representation attention mechanism to directly extract features from all patches of each WSI. Specifically, we first divide each WSI into bag-, patch- and cell-level images, and then assign the slide-level label to its corresponding bag-level images, so that WSI classification becomes a MIL problem. Additionally, we design a novel multi-scale representation attention mechanism, and embed it into a two-branch deep network to simultaneously mine the bag with a correct label, the significant patches and their cell-level information. Extensive experiments demonstrate the superior performance of the proposed framework over recent state-of-the-art methods, in term of classification accuracy and model interpretability. All source codes are released at: https://github.com/xhangchen/MRAN/.

ProMIL: A weakly supervised multiple instance learning for whole slide image classification based on class proxy

The histopathological analysis of a suspected region is critical for cancer diagnosis, treatment, and management. Histopathological diagnosis consists in analyzing the characteristics of the lesions using tissue sections stained with hematoxylin and eosin. Classification of digital tumor pathology images, called whole slide images (WSIs), is a great challenge since WSIs usually have huge resolutions while lacking localized annotations. Multiple instance learning (MIL) is a commonly used method applied to pathological image analysis. However, most MIL methods often focus only on the global representation of WSIs, ignoring whether the category labels play other roles in the model training besides being a supervision signal. In addition, feature confusion is also a problem that should be avoided for the analysis of WSIs with weakly supervised methods. To address these problems, we propose a novel algorithm of classifying WSI for cancer diagnosis. The proposed model, ProMIL, uses only slide-level labels rather than localized annotations for analysis. There are three innovations in this work. Firstly, we present the concept of class proxy which is the representation of the intrinsic feature of each category, and plays a key role in guiding the training of the model. Secondly, we design a novel WSI representation learning module that utilizes a multi-scale feature extraction strategy to represent each patch in a WSI and then aggregates these representations using an attention mechanism to encode the WSI. Thirdly, we design a metric-learning-based weakly supervised multiclass-classifier by measuring the similarity between each WSI embedding and class proxies. The proposed ProMIL can effectively alleviate the side effect of feature confusion, and carry intuitive interpretability and scalability. To evaluate the performance of ProMIL, we conduct a series of experiments on several datasets of WSIs with different types of cancer from open data sources. It can be observed from the experimental results that ProMIL outperforms most of the compared methods and achieves better performance on a various type of cancer image data for classification, thus suggesting the proposed method is suitable for classifying different categories of cancer rather than a specific kind of cancer. Therefore, it is expected to act as a general framework to be extended to more cancer diagnoses.

HistoMIL: A Python package for training multiple instance learning models on histopathology slides

Hematoxylin and eosin (H&E) stained slides are widely used in disease diagnosis. Remarkable advances in deep learning have made it possible to detect complex molecular patterns in these histopathology slides, suggesting automated approaches could help inform pathologists' decisions. Multiple instance learning (MIL) algorithms have shown promise in this context, outperforming transfer learning (TL) methods for various tasks, but their implementation and usage remains complex. We introduce HistoMIL, a Python package designed to streamline the implementation, training and inference process of MIL-based algorithms for computational pathologists and biomedical researchers. It integrates a self-supervised learning module for feature encoding, and a full pipeline encompassing TL and three MIL algorithms: ABMIL, DSMIL, and TransMIL. The PyTorch Lightning framework enables effortless customization and algorithm implementation. We illustrate HistoMIL's capabilities by building predictive models for 2,487 cancer hallmark genes on breast cancer histology slides, achieving AUROC performances of up to 85%.

Multiple instance learning from similarity-confidence bags

Multiple instance learning (MIL) is a classic weakly supervised learning approach, in which samples are grouped into bags that may contain varying numbers of instances. A bag is designated as positive if it contains at least one positive instance; otherwise, it is considered negative. Previous studies have consistently assumed that the bag labels are completely known. In fact, labeling every bag can be extremely challenging or even unfeasible due to the exorbitant expenses in terms of time and labor. Fortunately, it is much easier to obtain the similarity confidence, which represents the probability of two bags sharing the same label. How to employ it in MIL is worthy of study. Inspired by the above study, we present the first attempt to investigate MIL from similarity-confidence bags. Therefore, this paper proposes a new framework for training bag-level classifiers that adheres to the principle of empirical risk minimization. Moreover, we theoretically derive a generalization error bound to guarantee model convergence. Finally, we implement risk correction to mitigate potential over-fitting problem and provide theoretical consistency. Numerical experiments on eight datasets further validate the effectiveness of the proposed bag-level classifier.

Multiple instance learning to predict immune checkpoint blockade efficacy using neoantigen candidates

Multiple instance learning to predict immune checkpoint blockade efficacy using neoantigen candidates

Weakly-supervised anomaly detection with a Sub-Max strategy

We study weakly-supervised anomaly detection where only video-level “anomalous”/“normal” labels are available in training, while anomaly events should be temporally localized in testing. For this task, a commonly used framework is multiple instance learning (MIL), where clip instances are sampled from individual videos to form video-level bags. This sampling process arguably is a bottleneck of MIL. If too many instances are sampled, we not only encounter high computational overheads but also have many noisy instances in the bag. On the other hand, when too few instances are used, e.g., through enlarged grids, much background noise may be included in the anomaly instances. To resolve this dilemma, we propose a simple yet effective method named Sub-Max. In partitioned image regions, it identifies instances that are most probable candidates for anomaly events by selecting cuboids that have high optical flow magnitudes. We show that our method effectively brings down the computational cost of the baseline MIL and at the same time significantly filters out the influence of noise. Albeit simple, this strategy is shown to facilitate the learning of discriminative features and thus improve event classification and localization performance. For example, after annotating the event location ground truths of the UCF-Crime test set, we report very competitive accuracy compared with the state of the art on both frame-level and pixel-level metrics, corresponding to classification and localization, respectively.

A Cross-Domain Weakly Supervised Diabetic Retinopathy Lesion Identification Method Based on Multiple Instance Learning and Domain Adaptation.

Accurate identification of lesions and their use across different medical institutions are the foundation and key to the clinical application of automatic diabetic retinopathy (DR) detection. Existing detection or segmentation methods can achieve acceptable results in DR lesion identification, but they strongly rely on a large number of fine-grained annotations that are not easily accessible and suffer severe performance degradation in the cross-domain application. In this paper, we propose a cross-domain weakly supervised DR lesion identification method using only easily accessible coarse-grained lesion attribute labels. We first propose the novel lesion-patch multiple instance learning method (LpMIL), which leverages the lesion attribute label for patch-level supervision to complete weakly supervised lesion identification. Then, we design a semantic constraint adaptation method (LpSCA) that improves the lesion identification performance of our model in different domains with semantic constraint loss. Finally, we perform secondary annotation on the open-source dataset EyePACS, to obtain the largest fine-grained annotated dataset EyePACS-pixel, and validate the performance of our model on it. Extensive experimental results on the public dataset FGADR and our EyePACS-pixel demonstrate that compared with the existing detection and segmentation methods, the proposed method can identify lesions accurately and comprehensively, and obtain competitive results using only coarse-grained annotations.

Assessing the performance of fully supervised and weakly supervised learning in breast cancer histopathology

Fully supervised learning (FSL) and weakly supervised learning based on multiple instance learning (WSLMIL) have become two mainstream paradigms for performing computer-aided pathological diagnosis (CAPD). It is well known that the high-intensity annotation burden of FSL and the performance degradation due to poor training constraints of WSLMIL are stumbling blocks for clinical translation. Even more unfortunate is the lack of comprehensive experimental analysis to help researchers make content-specific trade-offs between FSL and WSLMIL. In this work, we systematically compare the performances of FSL and WSLMIL on lymph node metastasis in breast cancer using a publicly available dataset. By analyzing the results of 16 backbone networks in the FSL paradigm, we find that emerging networks based on transformer (PVTv2-B2) and multi-layer perceptron (CycleMLP-B3) are more advantageous for performing patch-level classification task than convolution-based structure (ResNet50); combining their output with morphological feature extraction can be better used to universally perform slide-level classification task. However, the slight improvement brought by the evolution of the backbone network may be overshadowed by the aggregation operation in 6 WSLMIL algorithms, whereas relying on the in-domain backbone network can achieve a stable and excellent prediction performance in both quantitative analysis and interpretability comparisons. All the experimental results ultimately illustrate that the combination of in-domain backbone network and emergent aggregation operation becomes an economical and efficient technical tool for CAPD, which can be regarded as a compromise between FSL and WSLMIL.

CNN-ViT Supported Weakly-Supervised Video Segment Level Anomaly Detection.

Video anomaly event detection (VAED) is one of the key technologies in computer vision for smart surveillance systems. With the advent of deep learning, contemporary advances in VAED have achieved substantial success. Recently, weakly supervised VAED (WVAED) has become a popular VAED technical route of research. WVAED methods do not depend on a supplementary self-supervised substitute task, yet they can assess anomaly scores straightway. However, the performance of WVAED methods depends on pretrained feature extractors. In this paper, we first address taking advantage of two pretrained feature extractors for CNN (e.g., C3D and I3D) and ViT (e.g., CLIP), for effectively extracting discerning representations. We then consider long-range and short-range temporal dependencies and put forward video snippets of interest by leveraging our proposed temporal self-attention network (TSAN). We design a multiple instance learning (MIL)-based generalized architecture named CNN-ViT-TSAN, by using CNN- and/or ViT-extracted features and TSAN to specify a series of models for the WVAED problem. Experimental results on publicly available popular crowd datasets demonstrated the effectiveness of our CNN-ViT-TSAN.

Registration‐enhanced multiple instance learning for cervical cancer whole slide image classification

AbstractHistopathological diagnosis is the golden standard for cancer diagnosis and grading. The past decade has witnessed great successes of vision‐based deep learning in computer‐aided cancer diagnosis. However, fully‐supervised methods require collecting large amounts of annotated data, which is a major challenge. In this research, we propose RMIL, a novel registration‐enhanced multiple instance learning pipeline, for cervical cancer biopsy grading of whole slide images as well as patch‐level regions. It requires only slide‐level annotations and a small amount of patch‐level annotations for training. To enhance it, an efficient registration method is implemented to generate more labeled patches with discriminative local features as data augmentations. In addition, attention mechanisms and self‐supervised learning are introduced to enhance feature extraction. Experiments showed that its classification accuracy is increased by 11.11%, compared to the vanilla multiple instance learning methods. The method has great potential for transferring to other scenarios.

Data-Efficient Computational Pathology Platform for Faster and Cheaper Breast Cancer Subtype Identifications: Development of a Deep Learning Model.

Breast cancer subtyping is a crucial step in determining therapeutic options, but the molecular examination based on immunohistochemical staining is expensive and time-consuming. Deep learning opens up the possibility to predict the subtypes based on the morphological information from hematoxylin and eosin staining, a much cheaper and faster alternative. However, training the predictive model conventionally requires a large number of histology images, which is challenging to collect by a single institute. We aimed to develop a data-efficient computational pathology platform, 3DHistoNet, which is capable of learning from z-stacked histology images to accurately predict breast cancer subtypes with a small sample size. We retrospectively examined 401 cases of patients with primary breast carcinoma diagnosed between 2018 and 2020 at the Department of Pathology, National Cancer Center, South Korea. Pathology slides of the patients with breast carcinoma were prepared according to the standard protocols. Age, gender, histologic grade, hormone receptor (estrogen receptor [ER], progesterone receptor [PR], and androgen receptor [AR]) status, erb-B2 receptor tyrosine kinase 2 (HER2) status, and Ki-67 index were evaluated by reviewing medical charts and pathological records. The area under the receiver operating characteristic curve and decision curve were analyzed to evaluate the performance of our 3DHistoNet platform for predicting the ER, PR, AR, HER2, and Ki67 subtype biomarkers with 5-fold cross-validation. We demonstrated that 3DHistoNet can predict all clinically important biomarkers (ER, PR, AR, HER2, and Ki67) with performance exceeding the conventional multiple instance learning models by a considerable margin (area under the receiver operating characteristic curve: 0.75-0.91 vs 0.67-0.8). We further showed that our z-stack histology scanning method can make up for insufficient training data sets without any additional cost incurred. Finally, 3DHistoNet offered an additional capability to generate attention maps that reveal correlations between Ki67 and histomorphological features, which renders the hematoxylin and eosin image in higher fidelity to the pathologist. Our stand-alone, data-efficient pathology platform that can both generate z-stacked images and predict key biomarkers is an appealing tool for breast cancer diagnosis. Its development would encourage morphology-based diagnosis, which is faster, cheaper, and less error-prone compared to the protein quantification method based on immunohistochemical staining.

DMIL-Transformer: Multiple Instance Learning via Integrating Morphological and Spatial Information for Lymph Node Metastasis Classification.

Automated classification of lymph node metastasis (LNM) plays an important role in the diagnosis and prognosis. However, it is very challenging to achieve satisfactory performance in LNM classification, because both the morphology and spatial distribution of tumor regions should be taken into account. To address this problem, this paper proposes a two-stage dMIL-Transformer framework, which integrates both the morphological and spatial information of the tumor regions based on the theory of multiple instance learning (MIL). In the first stage, a double Max-Min MIL (dMIL) strategy is devised to select the suspected top-K positive instances from each input histopathology image, which contains tens of thousands of patches (primarily negative). The dMIL strategy enables a better decision boundary for selecting the critical instances compared with other methods. In the second stage, a Transformer-based MIL aggregator is designed to integrate all the morphological and spatial information of the selected instances from the first stage. The self-attention mechanism is further employed to characterize the correlation between different instances and learn the bag-level representation for predicting the LNM category. The proposed dMIL-Transformer can effectively deal with the thorny classification in LNM with great visualization and interpretability. We conduct various experiments over three LNM datasets, and achieve 1.79%-7.50% performance improvement compared with other state-of-the-art methods.