Multiple Instance Learning Algorithms Research Articles

Simultaneous instance pooling and bag representation selection approach for multiple-instance learning (MIL) using vision transformer

In multiple-instance learning (MIL), the existing bag encoding and attention-based pooling approaches assume that the instances in the bag have no relationship among them. This assumption is unsuited, as the instances in the bags are rarely independent in diverse MIL applications. In contrast, the instance relationship assumption-based techniques incorporate the instance relationship information in the classification process. However, in MIL, the bag composition process is complicated, and it may be possible that instances in one bag are related and instances in another bag are not. In present MIL algorithms, this relationship assumption is not explicitly modeled. The learning algorithm is trained based on one of two relationship assumptions (whether instances in all bags have a relationship or not). Hence, it is essential to model the assumption of instance relationships in the bag classification process. This paper proposes a robust approach that generates vector representation for the bag for both assumptions and the representation selection process to determine whether to consider the instances related or unrelated in the bag classification process. This process helps to determine the essential bag representation vector for every individual bag. The proposed method utilizes attention pooling and vision transformer approaches to generate bag representation vectors. Later, the representation selection subnetwork determines the vector representation essential for bag classification in an end-to-end trainable manner. The generalization abilities of the proposed framework are demonstrated through extensive experiments on several benchmark datasets. The experiments demonstrate that the proposed approach outperforms other state-of-the-art MIL approaches in bag classification.

Cross-attention-based saliency inference for predicting cancer metastasis on whole slide images.

Although multiple instance learning (MIL) methods are widely used for automatic tumor detection on whole slide images (WSI), they suffer from the extreme class imbalance WSIs containing small tumors where the tumor may include only a few isolated cells. For early detection, it is important that MIL algorithms can identify small tumors. Existing studies have attempted to address this issue using attention-based architectures and instance selection-based methodologies but have not produced significant improvements. This paper proposes crossattention-based salient instance inference MIL (CASiiMIL), which involves a novel saliency-informed attention mechanism to identify small tumors (e.g., breast cancer lymph node micro-metastasis) on WSIs without needing any annotations. In addition to this new attention mechanism, we introduce a negative representation learning algorithm to facilitate the learning of saliencyinformed attention weights for improved sensitivity on tumor WSIs. The proposed model outperforms the state-ofthe-art MIL methods on two popular tumor metastasis detection datasets. The proposed approach demonstrates great cross-center generalizability, high accuracy in classifying WSIs with small tumor lesions, and excellent interpretability attributed to the saliency-informed attention weights. We expect that the proposed method will pave the way for training algorithms for early tumor detection on large datasets where acquiring fine-grained annotations is is not practical.

Dealing with imperfect data for invasive species detection using multispectral imagery

Detection and monitoring invasive species can provide valuable ecological information to guide management decisions. Multispectral imagery remote sensing may be an ideal tool to address this problem by providing accurate and affordable repeat imagery. However, developing training datasets for remote sensing imagery can be riddled with issues such as matching a training site to a single pixel in target imagery or mixed pixels caused by low invader densities. The multitarget multiple-instance spectral match filter (MTMI-SMF) algorithm accounts for such errors and has low computation costs. Here, we evaluated MTMI-SMF for invasive species detection with National Agriculture Imagery Program (NAIP), RapidEye, and Landsat imagery using Brazilian peppertree (Schinus terebinthifolia) in the Everglades National Park as a case study. MTMI-SMF detected Brazilian peppertree with an area under the curve (AUC) of 0.85 for Landsat, 0.82 for RapidEye, and 0.71 NAIP imagery. Invader classification was developed using a threshold calculated from MTMI-SMF image detection confidence values and compared to a random forest classification. MTMI-SMF and random forest classifications had similar overall accuracies with NAIP imagery performing the lowest (an average of 73.6% and 88.9%, respectively) and Landsat performing the highest (an average of 90.9% and 91.4%, respectively). However, due to the relatively rare nature of Brazilian peppertree across this landscape, the overall classification accuracy was not fully representative of detection capabilities. MTMI-SMF outperformed random forest for user accuracy and had comparable producer accuarcy across all three image sources. Classifications primarily relied on the brighter signature of Brazilian peppertree due to the limited number of spectral bands in the multispectral imagery. As a result, confusion was caused by heterogeneous vegetation communities or shrub habitats, which most likely could be resolved with the increased availability of satellite hyperspectral datasets. Our study indicates that MTMI-SMF, a machine-learning approach that allows flexibility in training data, can detect highly problematic invaders using multispectral imagery. This pipeline could be readily adapted to other invaders and ecosystems, as well as different remote sensing image sources, and has low computational requirements. Our study demonstrates that remote sensing technologies and multiple-instance learning algorithms can provide managers with critical tools for tackling the ever-growing, costly, and challenging problem of tracking invasive species' plant spread.

An end-to-end approach to combine attention feature extraction and Gaussian Process models for deep multiple instance learning in CT hemorrhage detection

Intracranial hemorrhage (ICH) is a serious life-threatening emergency caused by blood leakage inside the brain. Radiologists usually confirm the presence of ICH by analyzing computed tomography (CT) scans, so, developing an automated diagnosis system that can process this type of images has become an important research problem. One of the main challenges to apply AI algorithms in this setting is the lack of labeled data. To mitigate the labeling burden, Multiple Instance Learning (MIL) algorithms group instances into bags, relying solely on bag-level labels for model training. Due to their capacity to handle uncertainty and deliver accurate predictions, Gaussian Processes (GPs) stand out as promising classifiers for MIL problems. Recent research has also demonstrated the effectiveness of combining attention mechanisms with GPs for ICH detection. Nonetheless, existing methods have a notable limitation: they train the attention mechanism and the GP separately, resulting in suboptimal feature extraction for GP-based classification. In this study, we introduce an innovative end-to-end MIL model that concurrently trains the CNN backbone and attention mechanism along with the GP classifier. Our approach enhances the robustness and accuracy of bag predictions by optimizing feature extraction for GP-based classification. We validate our method experimentally by focusing on two ICH detection datasets. Our results reveal a significant performance advantage in terms of accuracy, F1-score, precision, and ROC-AUC score over existing MIL approaches, especially two-stage GP approaches. Additionally, we offer empirical insights into the functionality and effectiveness of our novel model.

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%.

RLogist: Fast Observation Strategy on Whole-Slide Images with Deep Reinforcement Learning

Whole-slide images (WSI) in computational pathology have high resolution with gigapixel size, but are generally with sparse regions of interest, which leads to weak diagnostic relevance and data inefficiency for each area in the slide. Most of the existing methods rely on a multiple instance learning framework that requires densely sampling local patches at high magnification. The limitation is evident in the application stage as the heavy computation for extracting patch-level features is inevitable. In this paper, we develop RLogist, a benchmarking deep reinforcement learning (DRL) method for fast observation strategy on WSIs. Imitating the diagnostic logic of human pathologists, our RL agent learns how to find regions of observation value and obtain representative features across multiple resolution levels, without having to analyze each part of the WSI at the high magnification. We benchmark our method on two whole-slide level classification tasks, including detection of metastases in WSIs of lymph node sections, and subtyping of lung cancer. Experimental results demonstrate that RLogist achieves competitive classification performance compared to typical multiple instance learning algorithms, while having a significantly short observation path. In addition, the observation path given by RLogist provides good decision-making interpretability, and its ability of reading path navigation can potentially be used by pathologists for educational/assistive purposes. Our code is available at: https://github.com/tencent-ailab/RLogist.

High-confidence AI-based biomarker profiling for H&E slides to optimize pathology workflow in lung cancer.

e21207 Background: In recent years AI methodologies have been increasingly adopted in the oncology space as part of the clinical decision process. Our previous research has shown accurate detection of actionable biomarkers based on AI algorithms applied to Hematoxylin & Eosin (H&E) scanned slides. While investigating a wide actionable biomarker panel, this profiling process can also flag patients with no actionable biomarkers. In Non-Small Cell Lung Cancer (NSCLC), EGFR, ALK, and ROS1 are the most commonly tested actionable alterations and the primary exclusion criteria for immunotherapy. Here we present a novel AI-based solution that can accurately and almost instantaneously detects patients without the main actionable biomarkers from the H&E image alone, thus helping prioritize the pathology lab workflow and navigate patients to a faster and more optimal diagnosis route. Methods: A retrospective cohort of 435 NSCLC patients was collected from the Sheba Medical Center pathology department. All cases (from primary and metastatic sites) with documented results for EGFR mutations, ALK, and ROS1-fusions were included. Molecular profiling was based on Oncomine Comprehensive, Dx target, or Solid tumor DNA assays, with 4 cases also supported by the Idylla EGFR mutation test. Deidentified digital whole slide images (WSIs) were run using Imagene’s AI classifier and compared to the officially reported results. The classifier was generated using self-supervised algorithm, combined with multiple instance learning algorithms applied using untagged and tagged FFPE H&E scan (40x or 20x magnification) from Imagene’s internal database (including TCGA research network slides). A categorical prediction (positive/negative) inferred on the validation set by the combined AI classifier model. Results: The combined AI model for detecting EGFR, ALK and ROS1 actionable alteration resulted in 127 patients (29.4% of the cohort) reported with no EGFR actionable mutation, ALK, or ROS1 fusions. Only two patients (1.57% of the negatively detected patients) were found to be false negative ( EGFR: p.G719D and p.L858R), demonstrating 98.43% accuracy. Conclusions: Image-based molecular detection can serve as an accurate and fast detection method to be used in the clinical setting for monitoring and prioritizing current pathology workflow. Specifically, it can detect a major fraction of the non-targetable patients and optimize their diagnostic flow.

P1164: IMAGE-BASED DETECTION OF HIGH-GRADE B CELL LYMPHOMAS (DH-L)

Background: Aggressive B cell Non-Hodgkin Lymphomas (B-NHL) are divided into two main categories: diffuse large B cell lymphoma (DLBCL) accounting for 90% of cases, and high-grade B-cell lymphoma (HGBL). Diagnosis of high-grade lymphoma with MYC and BCL2 and/or BCL6 rearrangements (double-hit lymphoma-DHL) is confirmed by fluorescence in situ hybridization (FISH) analysis, demonstrating c-MYC-rearrangement in combination with BCL-2 / BCL-6 rearrangement in lymphoma cells. Accurate and rapid diagnosis of DHL is obligatory, when considering more aggressive treatment regimens (other than R-CHOP), suggested in these patients. Aims: To establish a novel tool for diagnosing DLBCL and DHL, directly on the scanned Hematoxylin and Eosin (H&E) biopsy slides, by applying digital imaging technologies supported by machine learning algorithms. Methods: H&E whole slide images, prepared from biopsies obtained mainly from lymph nodes, but also from extra-nodal organs of patients with aggressive B cell lymphoma histology, were collected from the pathology department at Tel-Aviv Medical center (TASMC). Cases were randomly divided into a training (n=43) and a validation (n=35) set. On-the-fly augmentation was applied to images, together with advanced Convolutional Neural Network (CNN) analysis, to generate the aggressive B-NHL classifier (powered by Imagene-AI). Proprietary multiple instance learning (MIL) algorithms and training scheme, based on cases ranking, rather than absolute values, were applied. The classifier was validated through a testing set in a blinded study scheme, and results were compared to the FISH results for c-MYC and BCL2/6 rearrangements (Figure 1A). Results: The classifier training set was composed of 36 DLBCL NOS and 7 DHL cases. The aggressive B-NHL classifier was validated on a cohort of 35 cases, including 14 DHL cases (diagnosed by FISH) and 21 DLBCL cases. The model demonstrated 100% sensitivity and 90.48% specificity, with an accuracy rate of 94.29% and Area Under the Curve (AUC) of 0.99 (Figure 1B). Two cases were concluded as false positive, including one case that demonstrated rearrangements in substantial percentages of cells, just below the required positivity threshold. Image:Summary/Conclusion: Herein, we demonstrate for the first time a machine learning-based genomic testing solution for the detection and classification of aggressive B-NHL, based on H&E stained slide images. Implementation of this solution in clinical practice can support the diagnosis of DHL, which currently has both technical and financial challenges. Applying the proposed accessible and standardized AI-based method, can improve and facilitate the diagnosis of DHL, thereby directly improving patient care.

Deep multiple-instance learning for abnormal cell detection in cervical histopathology images

Cervical cancer is a disease of significant concern affecting women's health worldwide. Early detection of and treatment at the precancerous stage can help reduce mortality. High-grade cervical abnormalities and precancer are confirmed using microscopic analysis of cervical histopathology. However, manual analysis of cervical biopsy slides is time-consuming, needs expert pathologists, and suffers from reader variability errors. Prior work in the literature has suggested using automated image analysis algorithms for analyzing cervical histopathology images captured with the whole slide digital scanners (e.g., Aperio, Hamamatsu, etc.). However, whole-slide digital tissue scanners with good optical magnification and acceptable imaging quality are cost-prohibitive and difficult to acquire in low and middle-resource regions. Hence, the development of low-cost imaging systems and automated image analysis algorithms are of critical importance. Motivated by this, we conduct an experimental study to assess the feasibility of developing a low-cost diagnostic system with the H&E stained cervical tissue image analysis algorithm. In our imaging system, the image acquisition is performed by a smartphone affixing it on the top of a commonly available light microscope which magnifies the cervical tissues. The images are not captured in a constant optical magnification, and, unlike whole-slide scanners, our imaging system is unable to record the magnification. The images are mega-pixel images and are labeled based on the presence of abnormal cells. In our dataset, there are total 1331 (train: 846, validation: 116 test: 369) images. We formulate the classification task as a deep multiple instance learning problem and quantitatively evaluate the classification performance of four different types of multiple instance learning algorithms trained with five different architectures designed with varying instance sizes. Finally, we designed a sparse attention-based multiple instance learning framework that can produce a maximum of 84.55% classification accuracy on the test set.

CrowdDetective: Wisdom of the Crowds for Detecting Abnormalities in Medical Scans

Machine learning (ML) has great potential for early diagnosis of disease from medical scans, and at times, has even been shown to outperform experts. However, ML algorithms need large amounts of annotated data – scans with outlined abnormalities - for good performance. The time-consuming annotation process limits the progress of ML in this field. To address the annotation problem, multiple instance learning (MIL) algorithms were proposed, which learn from scans that have been diagnosed, but not annotated in detail. Unfortunately, these algorithms are not good enough at predicting where the abnormalities are located, which is important for diagnosis and prognosis of disease. This limits the application of these algorithms in research and in clinical practice. I propose to use the “wisdom of the crowds” –internet users without specific expertise – to improve the predictions of the algorithms. While the crowd does not have experience with medical imaging, recent studies and pilot data I collected show they can still provide useful information about the images, for example by saying whether images are visually similar or not. Such information has not been leveraged before in medical imaging applications. I will validate these methods on three challenging detection tasks in chest computed tomography, histopathology images, and endoscopy video. Understanding how the crowd can contribute to applications that typically require expert knowledge will allow harnessing the potential of large unannotated sets of data, training more reliable algorithms, and ultimately paving the way towards using ML algorithms in clinical practice.

Adaptively Converting Auxiliary Attributes and Textual Embedding for Video Captioning Based on BiLSTM

Automatically generating captions for videos faces a huge challenge since it is a cross-modal cross task that involves vision and texts. Most of the existing models generate the captioning words merely based on the video visual content features, ignoring the important underlying semantic information. The relationship between explicit semantics and hidden visual content is not holistically exploited, thus hardly describing fine-grained caption accurately from a global view. To better extract and integrate the semantic information, we propose a novel encoder-decoder framework of bi-directional long short-term memory with attention model and conversion gate (BiLSTM-CG), which transfers auxiliary attributes and then generates detailed captioning. Specifically, we extract semantic attributes from sliced frames in a multiple-instance learning (MIL) manner. MIL algorithms attempt to learn a classification function that can predict the labels of bags and/or instances in the visual content. In the encoding stage, we adopt 2D and 3D convolutional neural networks to encode video clips, and then feed the concatenate features into a BiLSTM. In decoding stage, we design a CG to adaptively fuse semantic attributes into hidden features at word level, and a CG can convert auxiliary attributes and textual embedding for video captioning. Furthermore, the CG has an ability to automatically decide the optimal time stamp to capture the explicit semantic or rely on the hidden states of the language model to generate the next word. Extensive experiments conducted on the MSR-VTT and MSVD video captioning datasets demonstrate the effectiveness of our method compared with state-of-the-art approaches.

Root identification in minirhizotron imagery with multiple instance learning

In this paper, multiple instance learning (MIL) algorithms to automatically perform root detection and segmentation in minirhizotron imagery using only image-level labels are proposed. Root and soil characteristics vary from location to location, thus, supervised machine learning approaches that are trained with local data provide the best ability to identify and segment roots in minirhizotron imagery. However, labeling roots for training data (or otherwise) is an extremely tedious and time-consuming task. This paper aims to address this problem by labeling data at the image level (rather than the individual root or root pixel level) and train algorithms to perform individual root pixel level segmentation using MIL strategies. Three MIL methods (multiple instance adaptive cosine coherence estimator, multiple instance support vector machine, multiple instance learning with randomized trees) were applied to root detection and compared to non-MIL approches. The results show that MIL methods improve root segmentation in challenging minirhizotron imagery and reduce the labeling burden. In our results, multiple instance support vector machine outperformed other methods. The multiple instance adaptive cosine coherence estimator algorithm was a close second with an added advantage that it learned an interpretable root signature which identified the traits used to distinguish roots from soil and did not require parameter selection.

Multiple-instance learning via multiple-point concept based instance selection

Multiple-instance learning (MIL) is a kind of weakly supervised learning where a single label is assigned to a bag of instances. To solve MIL problems, researchers have presented an effective embedding based framework that projects bags into a new feature space, which is constructed from some selected instances that can represent target concepts to some extent. Most previous studies use single-point concepts for the instance selection, where every possible concept is represented by only a single point (i.e., instance). However, multiple points may be more powerful for the same concept than a single. In this paper, we propose the notion of multiple-point concept, jointly represented by a group of similar points, and then build an iterative instance-selection method for MIL upon Multiple-Point Concepts. The proposed algorithm is thus named MILMPC, and its main difference from other MIL algorithms is selecting instances via multiple-point concept rather than single-point concept. The experimental results on five data sets have validated the convergence of the iterative instance-selection method, and the generality of the resulting MIL model in that it performs consistently well under three different kinds of relevance evaluation criteria (used to measure the relevance of a candidate concept to the target). Furthermore, compared to other MIL algorithms, the proposed model has been demonstrated not only suitable for common MIL problems, but more suitable for hybrid problems.

Visualization multi-instance data sets [Visualización de conjuntos de datos de múltiples instancias

In pattern recognition, multiple-instance learning algorithms have gained importance since they avoid that the user must delimit, the images individually in order to recognize the objects. This is an advantage over traditional learning algorithms since these considerably reduce the time required to prepare the data set. However, a disadvantage is that the resulting data sets are often complex, making it difficult to visualize them using traditional information visualization techniques. Thus, this work proposes a tool for the visualization and analysis of data sets of the multi-instance learning paradigm. The visualization proposal was evaluated using the expert criteria. In addition, different tests were carried out that show that a correct visualization can help to make decisions about the data set to improve the classification precision. © 2020, Associacao Iberica de Sistemas e Tecnologias de Informacao. All rights reserved.

Multiple-Instance Learning Approach via Bayesian Extreme Learning Machine

Multiple-instance learning (MIL) can solve supervised learning tasks, where only a bag of multiple instances is labeled, instead of a single instance. It is considerably important to develop effective and efficient MIL algorithms, because real-world datasets usually contain large instances. Known for its good generalization performance, MIL based on extreme learning machines (ELM-MIL) has proven to be more efficient than several typical MIL classification methods. ELM-MIL selects the most qualified instances from each bag through a single hidden layer feedforward network (SLFN) and trains modified ELM models to update the output weights. This learning approach often performs susceptible to the number of hidden nodes and can easily suffer from over-fitting problem. Using Bayesian inferences, this study introduces a Bayesian ELM (BELM)-based MIL algorithm (BELM-MIL) to address MIL classification problems. First, weight self-learning method based on a Bayesian network is applied to determine the weights of instance features. The most qualified instances are then selected from each bag to represent the bag. Second, BELM can improve the classification model via regularization of automatic estimations to reduce possible over-fitting during the calibration process. Experiments and comparisons are conducted with several competing algorithms on Musk datasets, images datasets, and inductive logic programming datasets. Superior classification accuracy and performance are demonstrated by BELM-MIL.

Texture analysis and multiple-instance learning for the classification of malignant lymphomas

Background and objectivesMalignant lymphomas are cancers of the immune system and are characterized by enlarged lymph nodes that typically spread across many different sites. Many different histological subtypes exist, whose diagnosis is typically based on sampling (biopsy) of a single tumor site, whereas total body examinations with computed tomography and positron emission tomography, though not diagnostic, are able to provide a comprehensive picture of the patient. In this work, we exploit a data-driven approach based on multiple-instance learning algorithms and texture analysis features extracted from positron emission tomography, to predict differential diagnosis of the main malignant lymphomas subtypes. MethodsWe exploit a multiple-instance learning setting where support vector machines and random forests are used as classifiers both at the level of single VOIs (instances) and at the level of patients (bags). We present results on two datasets comprising patients that suffer from four different types of malignant lymphomas, namely diffuse large B cell lymphoma, follicular lymphoma, Hodgkin’s lymphoma, and mantle cell lymphoma. ResultsDespite the complexity of the task, experimental results show that, with sufficient data samples, some cancer subtypes, such as the Hodgkin’s lymphoma, can be identified from texture information: in particular, we achieve a 97.0% of sensitivity (recall) and a 94.1% of predictive positive value (precision) on a dataset that consists in 60 patients. ConclusionsThe presented study indicates that texture analysis features extracted from positron emission tomography, combined with multiple-instance machine learning algorithms, can be discriminating for different malignant lymphomas subtypes.

MILDMS: Multiple Instance Learning via DD Constraint and Multiple Part Similarity

As a subject area of symmetry, multiple instance learning (MIL) is a special form of a weakly supervised learning problem where the label is related to the bag, not the instances contained in it. The difficulty of MIL lies in the incomplete label information of instances. To resolve this problem, in this paper, we propose a novel diverse density (DD) and multiple part similarity combination method for multiple instance learning, named MILDMS. First, we model the target concepts optimization with a DD function constraint on positive and negative instance space, which can greatly improve the robustness to label noise problem. Next, we combine the positive and negative instances in the bag (generated by hand-crafted and convolutional neural network features) with multiple part similarities to construct an MIL kernel. We evaluate the proposed approach on the MUSK dataset, whose results MUSK1 (91.9%) and MUSK2 (92.2%) show our method is comparable to other MIL algorithms. To further demonstrate generality, we also present experimental results on the PASCAL VOC 2007 and 2012 (46.5% and 42.2%) and COREL (78.6%) that significantly outperforms the state-of-the-art algorithms including deep MIL and other non-deep MIL algorithms.

Incremental learning of concept drift in Multiple Instance Learning for industrial visual inspection

Most Multiple Instance Learning (MIL) algorithms are designed with the assumption that the target concept is stationary in time, i.e. it is drawn from a stationary unknown distribution. However, in real industrial applications, like automatic visual inspection where defects may evolve, MIL has to deal with changing target concepts whose statistical characteristics may vary over time. Despite this fact, there is little discussion about how to learn from data in non-stationary environments (or data with concept drift) using multiple instance learners. In this work, an incremental MIL algorithm is proposed in order to learn non-stationary and recurrent target concepts in industrial visual inspection applications. Experiments on both synthetic and real-world datasets are conducted to test the performance of the proposed approach. Real-world datasets come from the automatic visual inspection task in industry. The experimental results show that the proposed approach is able to handle changing target concepts over time.

Anomaly classification in digital mammography based on multiple‐instance learning

Cancer tissues in mammography images exhibit abnormal regions; it is of great clinical importance to label a mammography image as having cancerous regions or not, perform the corresponding image segmentation. However, the detailed annotation of the cancer region is often an ambiguous and challenging task. The authors describe a fully automatic computer-aided detection and diagnosis (CAD) system to detect and classify breast cancer as malignant or benign, by using mammography and building on the multiple-instance learning (MIL) algorithms, which has been confirmed beneficial for radiologist decision sustenance. Traditional learning methods require great effort to annotate the training data by costly manual labelling and specialised computational models to detect these annotations during the test. The proposed CAD system simultaneously performs pixel-level segmentation (suspicious versus normal tissue) and image-level classification (benign versus malignant image). The set-up of the proposed system is in order: automatically segmented regions of interest (ROIs). Then, features derived from ROIs detected such as textural features and shape features are selected and extracted from each region and combined them to classify ROIs as `benign' or `malignant', by implementing MIL algorithms. Experimental results demonstrate the efficiency and robustness of the proposed CAD system compared with previous work in the literature.

Multiple instance learning: A survey of problem characteristics and applications

Multiple instance learning (MIL) is a form of weakly supervised learning where training instances are arranged in sets, called bags, and a label is provided for the entire bag. This formulation is gaining interest because it naturally fits various problems and allows to leverage weakly labeled data. Consequently, it has been used in diverse application fields such as computer vision and document classification. However, learning from bags raises important challenges that are unique to MIL. This paper provides a comprehensive survey of the characteristics which define and differentiate the types of MIL problems. Until now, these problem characteristics have not been formally identified and described. As a result, the variations in performance of MIL algorithms from one data set to another are difficult to explain. In this paper, MIL problem characteristics are grouped into four broad categories: the composition of the bags, the types of data distribution, the ambiguity of instance labels, and the task to be performed. Methods specialized to address each category are reviewed. Then, the extent to which these characteristics manifest themselves in key MIL application areas are described. Finally, experiments are conducted to compare the performance of 16 state-of-the-art MIL methods on selected problem characteristics. This paper provides insight on how the problem characteristics affect MIL algorithms, recommendations for future benchmarking and promising avenues for research. Code is available on-line at https://github.com/macarbonneau/MILSurvey.