IMPORTANCE Mammography screening currently relies on subjective human interpretation. Artificial intelligence (AI) advances could be used to increase mammography screening accuracy by reducing missed cancers and false positives. OBJECTIVE To evaluate whether AI can overcome human mammography interpretation limitations with a rigorous, unbiased evaluation of machine learning algorithms. DESIGN, SETTING, AND PARTICIPANTS In this diagnostic accuracy study conducted between September 2016 and November 2017, an international, crowdsourced challenge was hosted to foster AI algorithm development focused on interpreting screening mammography. More than 1100 participants comprising 126 teams from 44 countries participated. Analysis began November 18, 2016. MAIN OUTCOMES AND MEASUREMENTS Algorithms used images alone (challenge 1) or combined images, previous examinations (if available), and clinical and demographic risk factor data (challenge 2) and output a score that translated to cancer yes/no within 12 months. Algorithm accuracy for breast cancer detection was evaluated using area under the curve and algorithm specificity compared with radiologists' specificity with radiologists' sensitivity set at 85.9% (United States) and 83.9% (Sweden). An ensemble method aggregating top-performing AI algorithms and radiologists' recall assessment was developed and evaluated. RESULTS Overall, 144 231 screening mammograms from 85 580 US women (952 cancer positive Յ12 months from screening) were used for algorithm training and validation. A second independent validation cohort included 166 578 examinations from 68 008 Swedish women (780 cancer positive). The top-performing algorithm achieved an area under the curve of 0.858 (United States) and 0.903 (Sweden) and 66.2% (United States) and 81.2% (Sweden) specificity at the radiologists' sensitivity, lower than community-practice radiologists' specificity of 90.5% (United States) and 98.5% (Sweden). Combining top-performing algorithms and US radiologist assessments resulted in a higher area under the curve of 0.942 and achieved a significantly improved specificity (92.0%) at the same sensitivity. CONCLUSIONS AND RELEVANCE While no single AI algorithm outperformed radiologists, an ensemble of AI algorithms combined with radiologist assessment in a single-reader screening environment improved overall accuracy. This study underscores the potential of using machine (continued)

Several computer aided diagnosis (CAD) systems have been developed for mammography. They are widely used in certain countries such as the U.S. where mammography studies are conducted more frequently; however, they are not yet globally employed for clinical use due to their inconsistent performance, which can be attributed to their reliance on hand-crafted features. It is difficult to use hand-crafted features for mammogram images that vary due to factors such as the breast density of patients and differences in imaging devices. To address these problems, several studies have leveraged a deep convolutional neural network that does not require hand-crafted features. Among the recent object detectors, RetinaNet is particularly promising as it is a simpler one-stage object detector that is fast and efficient while achieving state-of-the-art performance. RetinaNet has been proven to perform conventional object detection tasks but has not been tested on detecting masses in mammograms. Thus, we propose a mass detection model based on RetinaNet. To validate its performance in diverse use cases, we construct several experimental setups using the public dataset INbreast and the in-house dataset GURO. In addition to training and testing on the same dataset (i.e., training and testing on INbreast), we evaluate our mass detection model in setups using additional training data (i.e., training on INbreast + GURO and testing on INbreast). We also evaluate our model in setups using pre-trained weights (i.e., using weights pre-trained on GURO, training and testing on INbreast). In all the experiments, our mass detection model achieves comparable or better performance than more complex state-of-the-art models including the two-stage object detector. Also, the results show that using the weights pre-trained on datasets achieves similar performance as directly using datasets in the training phase. Therefore, we make our mass detection model’s weights pre-trained on both GURO and INbreast publicly available. We expect that researchers who train RetinaNet on their in-house dataset for the mass detection task can use our pre-trained weights to leverage the features extracted from the datasets.

BackgroundAccurately detecting and examining lung nodules early is key in diagnosing lung cancers and thus one of the best ways to prevent lung cancer deaths. Radiologists spend countless hours detecting small spherical-shaped nodules in computed tomography (CT) images. In addition, even after detecting nodule candidates, a considerable amount of effort and time is required for them to determine whether they are real nodules. The aim of this paper is to introduce a high performance nodule classification method that uses three dimensional deep convolutional neural networks (DCNNs) and an ensemble method to distinguish nodules between non-nodules.MethodsIn this paper, we use a three dimensional deep convolutional neural network (3D DCNN) with shortcut connections and a 3D DCNN with dense connections for lung nodule classification. The shortcut connections and dense connections successfully alleviate the gradient vanishing problem by allowing the gradient to pass quickly and directly. Connections help deep structured networks to obtain general as well as distinctive features of lung nodules. Moreover, we increased the dimension of DCNNs from two to three to capture 3D features. Compared with shallow 3D CNNs used in previous studies, deep 3D CNNs more effectively capture the features of spherical-shaped nodules. In addition, we use an alternative ensemble method called the checkpoint ensemble method to boost performance.ResultsThe performance of our nodule classification method is compared with that of the state-of-the-art methods which were used in the LUng Nodule Analysis 2016 Challenge. Our method achieves higher competition performance metric (CPM) scores than the state-of-the-art methods using deep learning. In the experimental setup ESB-ALL, the 3D DCNN with shortcut connections and the 3D DCNN with dense connections using the checkpoint ensemble method achieved the highest CPM score of 0.910.ConclusionThe result demonstrates that our method of using a 3D DCNN with shortcut connections, a 3D DCNN with dense connections, and the checkpoint ensemble method is effective for capturing 3D features of nodules and distinguishing nodules between non-nodules.

Background: Since nuclei segmentation in histopathology images can provide key information for identifying the presence or stage of a disease, the images need to be assessed carefully. However, color variation in histopathology images, and various structures of nuclei are two major obstacles in accurately segmenting and analyzing histopathology images. Several machine learning methods heavily rely on hand-crafted features which have limitations due to manual thresholding. Results: To obtain robust results, deep learning based methods have been proposed. Deep convolutional neural networks (DCNN) used for automatically extracting features from raw image data have been proven to achieve great performance. Inspired by such achievements, we propose a nuclei segmentation method based on DCNNs. To normalize the color of histopathology images, we use a deep convolutional Gaussian mixture color normalization model which is able to cluster pixels while considering the structures of nuclei. To segment nuclei, we use Mask R-CNN which achieves state-of-the-art object segmentation performance in the field of computer vision. In addition, we perform multiple inference as a post-processing step to boost segmentation performance. We evaluate our segmentation method on two different datasets. The first dataset consists of histopathology images of various organ while the other consists histopathology images of the same organ. Performance of our segmentation method is measured in various experimental setups at the object-level and the pixel-level. In addition, we compare the performance of our method with that of existing state-of-the-art methods. The experimental results show that our nuclei segmentation method outperforms the existing methods. Conclusions: We propose a nuclei segmentation method based on DCNNs for histopathology images. The proposed method which uses Mask R-CNN with color normalization and multiple inference post-processing provides robust nuclei segmentation results. Our method also can facilitate downstream nuclei morphological analyses as it provides high-quality features extracted from histopathology images.

Assessing the status of metastasis in sentinel lymph nodes (SLNs) by pathologists is an essential task for the accurate staging of breast cancer. However, histopathological evaluation of SLNs by a pathologist is not easy and is a tedious and time-consuming task. The purpose of this study is to review a challenge competition (HeLP 2018) to develop automated solutions for the classification of metastases in hematoxylin and eosin-stained frozen tissue sections of SLNs in breast cancer patients. Materials and Methods A total of 297 digital slides were obtained from frozen SLN sections, which include postneoadjuvant cases (n=144, 48.5%) in Asan Medical Center, South Korea. The slides were divided into training, development, and validation sets. All of the imaging datasets have been manually segmented by expert pathologists. A total of 10 participants were allowed to use the Kakao challenge platform for 6 weeks with two P40 GPUs. The algorithms were assessed in terms of the area under receiver operating characteristic curve (AUC). Results The top three teams showed 0.986, 0.985, and 0.945 AUCs for the development set and 0.805, 0.776, and 0.765 AUCs for the validation set. Micrometastatic tumors, neoadjuvant systemic therapy, invasive lobular carcinoma, and histologic grade 3 were associated with lower diagnostic accuracy. Conclusion In a challenge competition, accurate deep learning algorithms have been developed, which can be helpful in making frozen diagnosis of intraoperative SLN biopsy. Whether this approach has clinical utility will require evaluation in a clinical setting.