Deep Regional Metastases Segmentation for Patient-Level Lymph Node Status Classification

Generally, automatic diagnosis of the presence of metastases in lymph nodes has therapeutic implications for breast cancer patients. Detection and classification of breast cancer metastases have high clinical relevance, especially in whole-slide images of histological lymph node sections. Fast early detection leads to huge improvement of patient’s survival rate. However, currently pathologists mainly detect the metastases with microscopic assessments. This diagnosis procedure is extremely laborious and prone to inevitable missed diagnoses. Therefore, automated, accurate patient-level classification would hold great promise to reduce the pathologist’s workload while also reduce the subjectivity of diagnosis. In this paper, we provide a novel deep regional metastases segmentation (DRMS) framework for the patient-level lymph node status classification. First, a deep segmentation network (DSNet) is proposed to detect the regional metastases in patch-level. Then, we adopt the density-based spatial clustering of applications with noise (DBSCAN) to predict the whole metastases from individual slides. Finally, we determine patient-level pN-stages by aggregating each individual slide-level prediction. In combination with the above techniques, the framework can make better use of the multi-grained information in histological lymph node section of whole-slice images. Experiments on large-scale clinical datasets (e.g., CAMELYON17) demonstrate that our method delivers advanced performance and provides consistent and accurate metastasis detection in clinical trials.

. Example of a typical metastatic region. Whole-slide images (WSI) are generally stored in a multi-resolution pyramid structure. Each image in the pyramid is stored as a series of tiles to facilitate rapid retrieval of subregions of the image (https://camelyon17.grand-challenge. org/Background/). The regions of metastases have very similar appearances to the normal regions of scanned sections.
important prognostic factors in breast cancer [1], [2]. The prognosis is poor when cancer spreads to the lymph nodes. Therefore, lymph nodes are surgically removed and examined under endoscope. However, the diagnostic procedures are tedious and time-consuming. Most importantly, small metastases are very difficult to detect and are sometimes missed.
Currently, the UICC TNM classification system is an internationally recognized means of classifying the extent of cancer spread in patients with solid tumors. It is one of the most important tools to help clinicians select suitable treatment options and obtain prognostic indications. In breast cancer, UICC TNM staging takes into account the size of the tumor (T-stage), whether the cancer has spread to the regional lymph nodes (N-stage), and whether the tumor has metastasised to other parts of the body (M-stage). Since the histological assessment of lymph node metastases is an essential part of UICC TNM classification, in this paper we will focus on the pathologic N-stage (pN-stage).
Automatically detecting the regional metastases, especially from the whole-slide images, is a very challenging task due to the low contrast and similar structures between metastasised regions and various adjacent tissues or organs, as shown in Fig. 1. Moreover, the interference from other tumors makes the metastasised regions even more difficult to be separated from the background. In some cases, the metastasised regions are extremely small, which further complicates the task.
In this paper, we propose a deep regional metastases segmentation (DRMS) framework to predict pN-stage scores from patient's whole slide histopathology images. The proposed framework consists of four key modules: a region of interest (ROI) extraction module, a patch-level regional metastases localization module, a slide-level metastasis detection module and a patient-level lymph node classification module. First, the ROI extraction module generates candidate tissue regions from whole slide images (WSIs). Second, the patch-level regional metastases localization module adopts a deep segmentation network (DSNet) to predict cancer metastases within extracted ROIs. Third, the slide-level metastasis detection module adopts the DBSCAN to fuse the predicted scores extracted from ROIs to build a slide-level lymph node prediction [3]. Finally, patientlevel pN-stage is determined by aggregating slide-level predictions.
In summary, our contributions are four folds: 1. We propose a DRMS framework to predict pathologic N-stage (pN-stage) from patient's whole slide histopathology images.
2. We propose a multi-grained procedure for detecting cancer metastases on giga-pixel pathology images. With multigrained analysis, the proposed method can capture rich information from pathology images.
3. We propose a deep segmentation network (DSNet) for patch-level metastases localization. The proposed network can aggregate the global-to-local context information of each patch.
4. Experiments on public benchmarks demonstrate that our approach achieves superior performance and obtains accurate results on metastases detection.
The rest of this paper is organized as follows. In Sect. 2, we give an overview of classical and deep learning methods for lymph node metastases detection, status classification and tumor segmentation. Then we introduce the proposed approach in Sect. 3. In Sect. 4, we evaluate and analyze the proposed method by extensive experiments and comparisons with other methods. Finally, we provide the conclusion and future work in Sect. 5.

II. RELATED WORKS
Our work aims to introduce deep learning-based segmentation techniques to detect and locate the metastases regions, and further predict the lymph node status. Some concerning works related to breast cancer metastases detection, lymph node status classification and deep learning-based tumor segmentation methods will be reviewed. For more details, we refer readers to comprehensive surveys [4], [5].

A. BREAST CANCER METASTASES DETECTION
Breast cancer has been studied in few decades due to its constructive effect for human health. Staging the breast cancer is a central component of treatment and management, which involves the microscopic examination of lymph nodes adjacent to the breast for evidence that the cancer has spread or metastasized [6], [7]. This process requires highly skilled pathologists and is quite time-consuming and error-prone, especially for lymph nodes with no or small tumors. Computer aided detection (CAD) of lymph node metastasis could improve the sensitivity, speed, and consistency of metastasis detection. However, only recently metastases detection has been introduced to the breast cancer diagnosis with power models. For example, Bejnordi first proposed a fully automated detection of DCIS in breast histopathology images [8]. Then, Bejnordi et al. proposed a context-aware convolutional neural network (CNN) for classification of breast carcinomas [9]. Bulten and Litjens [10] proposed an unsupervised prostate cancer detection on H&E images with deep auto-encoders. Pinckaers and Litjens [11] trained CNNs with megapixel images for cancer localization. Though effectiveness, previous methods mainly consider low-resolution histopathology images. In modern hospitals, the scanning and digitization of whole histology slides are adopted to improve the tissue analysis for efficient cancer diagnosis and prognosis. How to handle the WSI for cancer metastases detection is still a challenging task.

B. LYMPH NODE STATUS CLASSIFICATION
For the lymph node status classification, well trained histopathologists can provide accurate classification results. However, due to the huge size of WSIs and the great number of potential cancer cases, fully automatic solutions for lymph node status classification is highly desirable. In the recent past, many related techniques have been proposed. The most impactful thing is organizing the challenges of CAMELYON16 and CAMELYON17. Several promising works have applied deep learning to lymph node status classification. For the CAMELYON16 challenge, the winner has shown an expressive performance for per slide and overall slide-level classification [12]. In their work, a modified GoogleNet is trained with a pre-sampled set of image patches to predict the coarse location of cancer metastases, then a random forest classifier is used to predict the slide label [13]. The CAMELYON organizers also trained CNNs on smaller datasets to detect breast cancer in lymph nodes and prostate cancer [14]. With new modules, several methods show state of the art results in lymph node status classification. However, they generally introduce a variety of image analysis tasks, resulting to high computation and memory requirements. Different from previous works, we observe that the granularity and variability of WSIs can help to improve the performances of lymph node status classification. Thus, we develop an integrated framework for WSI analysis. The proposed framework can make better use of the multi-grained information in histological lymph node section of whole-slice images.

C. HISTOPATHOLOGICAL TUMOR SEGMENTATION
Tumor segmentation in histopathology image can be coarsely categorized into hand-crafted feature-based methods and deep-learned feature-based methods. In the past decades, many hand-crafted features-based methods have been proposed to solve the tumor segmentation problems of histopathology image. For example, threshold-based methods, such as Otsu method classify foreground and background based on whether the intensity value is above a threshold [15]. Its segmentation performance has obvious bottlenecks for complex histopathology images. Due to the limited representation ability of the hand-crafted feature, the segmentation performance has not been effectively improved.
With the rapid development of deep learning technologies and large-scale datasets, several outstanding methods have been proposed in tumor segmentation to improve segmenta-tion performance. Ronneberger et al. presented the U-shape network architecture consisting of a contracting path and a symmetric expanding path for biomedical image segmentation [16]. In general, U-Net has proven to be one of the most popular architectures for medical image segmentation tasks. Zhou et al. presented a novel U-Net++ architecture which combined a series of nested dense skip pathways [17]. Its segmentation performance is superior to other methods in medical image segmentation tasks. Qaiser et al. utilized the concept of persistent homology profiles (PHPs) for adaptive tumor segmentation [18]. Because the various appearances of tumors, hard samples are also one of the common challenges in tumor segmentation. The works in Lin et al. and Wang et al. adopted hard negative mining to obtain better performance in different tumor segmentation [12], [19], [20]. In this work, we propose a novel network for improving the ability of localizing the cancer regions. Besides, we also introduce a new multi-grained procedure for detecting cancer metastases on giga-pixel pathology images. Fig. 2 illustrates the overall procedure of our proposed framework. The whole framework consists of four key modules: a region of interest (ROI) extraction module, a patch-level regional metastases location module, a slide-level metastasis detection module and a patient-level lymph node classification module. We start with identifying the ROIs and removing non-information regions on the WSIs. Subsequently, we feed image patches extracted from the ROIs into our deep segmentation network (DSNet) for coarse metastases prediction. After that, we adopt DBSCAN to fuse the patch-level prediction and achieve the slide-level lymph node metastasis prediction [3]. Meanwhile, we extract corresponding features from the predicted metastases area. We finally train a XGBoost [21] classifier with these features to predict the pN-stage of each patient. In the following section, we will describe the key components in detail.

A. ROI EXTRACTION
For histopathology images, a typical whole slide image is approximately 200, 000 × 100.000 pixels on the highest resolution level. If we deal with such large images directly, enormous computation is required because of the huge size of the slide. Besides, most of the regions have no useful information for metastases detection. Thus, directly dividing WSIs into patches is not appropriate for model training, as shown in Fig. 3. The Otsu threshold [15] method is widely used in recent studies for ROI extractions from the WSI because of its simple calculation and the segmentation is not affected by image brightness and contrast. In this work, to speed up the processing, we also adopt Otsu's adaptive thresholding technique on the low-resolution version of the WSIs to remove most of the non-tissue background quickly. Following [22], we convert RGB to gray from 32-times down-sampled WSI and then extract tissue regions by Otsu  threshold method. We observed that metastasis regions are usually located at the edge of the tissue regions. Therefore, careful tissue region extraction method is needed. After the operation is done, holes and small points, which will influence feature extraction operation, still exist. Following previous works, morphological algorithms are operated to fill holes and clear isolated points. Then, we generate the patches through random grid sampling. Examples of the resulting patches are shown in Fig. 4. We can see that although most examples have texture patterns, it is difficult to distinguish between normal or metastatic patches.

B. PATCH-LEVEL METASTASES LOCALIZATION
The patch-level localization aims to predict the metastases regions at coarse granularities. To achieve this goal, we opt to a deep segmentation network, inspired by [23], as shown in Fig. 5. The DSNet takes in the patch images, extract multi-level features with dilated ResNet-101 [24]. (Dilation size = 2 at the last residual module), resulting in 1/16 resolution of input images and extracts locally weighted, multiscale contextual information with the proposed pyramidal attentional ASPP (PA-ASPP).  The network's ability to understand and utilize contextual information is critical to achieving accurate image segmentation. Increasing the network receptive field by consecutive down sampling is a straightforward way to capture contextual clues for dense prediction. However, the receptive field is restricted by the number of down-sampling operations performed, as performing down-sampling can cause spatial information loss. Inspired by the methods of SPP [25] and ASPP [23], [26] and the approach used by radiologists to locate regional metastases by observing neighboring contexts, we propose the PA-ASPP for extracting locally weighted, multiscale contextual information. The PA-ASPP feeds the input into two branches, one is to the pyramid feature aggregation (top branch) for providing pixel-level attention and the other one is to the ASPP (bottom branch) for providing rich multi-scale features and large receptive fields. The pyramid feature aggregation encodes the multi-level features from three different scales like U-Net [16]. This feature fusion incorporates neighboring scales of context features more precisely to produce better pixel-level attention.
The obtained attention is up-sampled to the same size of output of ASPP and used to weight the multi-scale features in pixel-level for selecting contextual information. The ASPP enriches contextual information by concatenating features obtained from one 1 × 1 convolution, three 3 × 3 convolutions with dilation rates equal to (6,12,18), and one global average pooling layer and then passing through another 1 × 1 convolution to modify channel dimensions of multi-scale features before multiplying the attention maps. Finally, we use a scale-aware selection (SAS) module to locally weight multiscale contextual features. More specifically, we first use a squeeze-and-excitation module to select the key channels from multi-scale features [27]. Then, the selected features are highlighted pixel-wisely by attention maps from Pyramid Attention Module (PAM). The SAS can capture important contexts of ASPP for the accurate and consistent dense prediction.
To train the proposed network, we adopt the input resolution of 512 × 512 to capture more detailed information. To increase the variation in the training set, we perform extensive data augmentation including random flipping, scaling, rotation over angles between 0 and 360, and cropping. Besides, our model is trained with hard negative mining to improve the discrimination. It means that for each epoch, model inferences whole patches of the slide and chooses patches whose intersection over union (IOU) with the annotated mask are less than 0.95 as the training set.

C. SLIDE-LEVEL METASTASES DETECTION
To capture the slide-level information, we should integrate the patch information to improve the spatial consistency. Following the work in [28], we introduce the DBSCAN algorithm to group together small metastases areas which were in close proximity.

D. SPATIENT-LEVEL CLASSIFICATION
After slide-level metastases detection, we can realize the patient-level classification based on the corresponding metastases regions. More specifically, each whole slide probability map is first converted into a feature vector after being regulated with 5 threshold values, i.e., 0.5, 0.6, 0.7, 0.8, 0.9, which is then used to train a pN-stage classifier. According to the morphological and geometrical information, we extract 13 types of features from each probability map (65 features in total). The task is to determine a pN-stage for each patient in the test dataset. To compose a pN-stage, the number of positive lymph nodes (i.e., nodes with a metastasis) need to be counted. According to the size of the tumor, there are three types of positive lymph nodes in clinic: Macro-metastases: Metastases greater than 2.0 mm. Micro-metastases: Metastases greater than 0.2 mm or more than 200 cells, but smaller than 2.0 mm.
Isolated Tumor Cell (ITC): Single tumor cells or a cluster of tumor cells smaller than 0.2 mm or less than 200 cells.
Although lymph nodes containing only ITCs are not counted as positive lymph nodes, if no macro-metastases or micro-metastases are found in the patient's lymph nodes, the pathologist must report ITC.
In this work, we follow previous works and adopt a simplified version of the pN-staging system in breast cancer. The task is to automatically determine the following applied pN-stages for each patient: -pN0: No micro-metastases or macro-metastases or ITCs found.
-pN1: Metastases found in 1-3 lymph nodes, of which at least one is a macro-metastasis.
-pN2: Metastases found in 4-9 lymph nodes, of which at least one is a macro-metastasis.
After feature vectors are extracted, we use XGBoost to classify lymph nodes into four classes (Normal, ITC, Micro, Macro). Since the total number of WSIs is rather small, it is important to prevent overfitting. Thus, K-fold cross validation and other hyper-parameters, such as subsample or max depth, are adopted for preventing overfitting. Firstly, grid-search algorithm is used to search for the best hyper-parameters. Then, the XGBoost model is trained with ten-fold cross validation. Finally, the best model is selected and the 100 patients in the CAMELYON17 test set are predicted. The patient's pN-stage is graded according to the given rules.

E. LOSS FUNCTION
Tumor regions cover a very small proportion of pixels in WSIs, thereby leading to class imbalance. This issue was circumvented by training the network to minimize a hybrid loss function. In this work, the hybrid cost function comprised of a class-weighted cross-entropy loss and a loss function based on the Dice overlap coefficient. The Dice coefficient is an overlap metric used for assessing the quality of segmentation maps. The Dice loss is a differentiable function that approximates to Dice coefficient, it uses the predicted posterior probability map and ground truth binary image, as defined in Eq. 1. The weighted cross-entropy loss is defined in Eq. 2. In the equations, p ln is the predicted posterior probability map, and g ln is the ground truth image.
The total loss is defined as a linear combination of the two loss components as defined in Eq. 3.
The proposed networks are trained by minimizing the total loss. α and β are empirically assigned to the individual loss components. In this work we set α = β = 0.5.

IV. EXPERIMENTS
In this section, we first introduce our experimental settings such as public benchmarks, training details, and evaluation VOLUME 9, 2021 metrics. Then we report the experimental results on public benchmarks. Finally, we analyze the key components with extensive ablation experiments.

A. EXPERIMENTAL SETTINGS 1) DATASETS
In this work, we evaluate our framework on CAMELYON17 dataset [29]. However, following previous works, we also used the WSIs from CAMELYON16 as the training dataset. The CAMELYON16 dataset contains 400 WSIs with region annotations for all its metastasis slides [30]. The WSIs are collected from two different medical centers. The CAMELYON17 dataset contains 1000 WSIs with 5 slides per patient: 500 slides for the train set, 500 slides for the test set. The WSIs are collected from five different medical centers. Typically, different sizes of breast cancer metastases are in sentinel lymph nodes, as shown in Fig. 6. Since the CAMELYON17 dataset provides only 50 slides with region annotations, we split 100 patients (500 slides) into 43 patients for the train set, 57 patients for the validation set for hyperparameter tuning. In detail, if a patient's any slide includes region annotation, we allocate that patient as a train set. For training of the patch-level localization, 400 WSIs from CAMELYON16 dataset and 160 WSIs (50 WSIs with region annotation and 110 negative WSIs) from CAMELYON17 train set are used. For training of the slide-level lymph node classifier, we use 285 WSIs (57 patients) from CAMELYON17 validation set.

2) EVALUATION METRICS
To evaluate the performance of different methods, we adopt the quadratic weighted Cohen's kappa as the evaluation metric [31]. Given n test patients and m categories (pN-stages), let n ij denote the number of patients with the pNi-stage that were categorized to pNj-stage. Let r i denote the total number of patients with the pNi-stage and s j the total number of patients categorized to the pNj-stage. Finally, let w ij denote the disagreement weight associated with the pN i and pN j . The weight matrix is The weight matrix is The mean observed degree of disagreement is The mean degree of disagreement expected by chance is The weighted kappa is then defined by The k w metric ranges from −1 to +1: a negative value indicates lower than chance agreement, zero indicates exact chance agreement, and a positive value indicates better than chance agreement. As pN-stages are ordinal, choosing a quadratic weighted kappa could penalize misclassification more severely when errors are more than one stage apart.

3) TRAINING DETAILS
During training and inference, we first perform the thresholding for localizing the ROIs, then extracted 512 × 512 patches from WSIs at the highest magnification level resolution. We adopt the ImageNet pre-trained ResNet101 (He et al. 2016) as the backbone with dilation convolutions. New layers are initialized with the MSRA method. The deep learning toolbox TensorFlow was used to train the model with a NVIDIA GTX TITAN X GPU [32]. We used the Adam optimization method with a learning rate 1e-4. The network was trained for 4 epochs with a batch size 32 [33]. For the post-processing, we regulated tumor probability heat map with 5 threshold values, i.e., 0.5, 0.6, 0.7, 0.8, 0.9. Given a heat map, we extracted 13 features including the major axis length of the tumor region (Feature1), maximum probability score (Feature2), average probability score (Feature3), a total area of the tumor region (Feature4), max, mean, variance of eccentricity (Feature5-7), orientation (Feature8-10) and solidity (Feature11-13) from the tumor regions. We built a XGBoost classifier to discriminate lymph node classes using extracted features. Finally, each patient's pN-stage was determined by the given rule together with the lymph node slide prediction result.

B. COMPARISONS WITH OTHER STATE-OF-THE-ART METHODS
Using a single segmentation model, our slide-level lymph node classification model and patient-level pN-stage prediction achieved 0.9351 accuracy and 0.9017 quadratic weighted kappa score, respectively. We trained additional models with different data augmentation. Finally, four models were ensembled by averaging probability heat map and reached 0.9410 slide-level accuracy and 0.9632 patient-level quadratic weighted kappa score. Tab. 1 shows the published results on CAMELYON17 testing dataset. Our best performing model gave a Cohen's kappa score of 0.9473, which is better than the second in Open Leaderboard (out of 120 valid submission entries). According to the brief description of the top-ranked model, it also used DeepLabv3+ as the basic framework for patch level segmentation. However, the  top-ranked model used the DenseNet with high generalization ability as the backbone of the Encoder. Our proposed DSNet+ adopted Resnet as the backbone and improve the DeepLabv3+ framework by introducing the PA and SAS attentions to enhance the model generalization ability. In the future, we will explore more advanced segmentation frameworks and backbone networks for the dedicated task. Fig. 7 shows an example for slide-level predictions. From the results, we can see that our method can achieve outstanding performance in different cases, including continuous regions (a), isolated regions (b) and low-contrast regions (c). In (a) the predicted whole tumor segmentation was similar to pathologist provided ground truth of whole tumor region and most of the samples in the dataset fell into this category.

C. ABLATIONS STUDIES
Comprehensive experiments are conducted to demonstrate the effectiveness of the proposed method. The 5-fold crossvalidation is applied to validate the stability and generalization of networks. It is worth mentioning that all models are trained following the same experimental settings.

1) EFFECTS OF THE PA-ASPP AND SAS
The proposed PA-ASPP is used to aggregate the global-tolocal context information. The PPM [25] and ASPP [23] also aggregate multi-scale contextual features for semantic labeling. To verify the effects of our PA-ASPP, we adopt the PSPNet [25] with the ResNet-101 backbone for a fair comparison. We replace the PPM of PSPNet with our PA-ASPP and ASPP modules. The 2-4 rows of Tab. 1 show the quantitative results. With our PA-ASSP module, the Cohen Kappa score improves about 3.3% and 1.6% over the PPM and ASPP, respectively. Visual comparisons of PPM, ASPP and our PA-ASPP are shown in Fig. 8. It can be observed that our segmentation results constantly contain more accurate and coherent structures. These results demonstrate that our proposed PA-ASPP is more effective than previous methods and can generate more coherent results. Besides, with the SAS, the model can adaptively capture the regions with different scales, which shows the effectiveness.

2) EFFECTS OF HARD EXAMPLE MINING
Previous patch extraction scheme from the whole-slide images is based on random uniform sampling. However, this leads to some hard examples being excluded from the training set which resulted in the model's poor performance on such regions of WSI. Therefore, we attempt to solve this issue by hard example mining the poorly performing regions in WSI and fine-tuning the trained model with this hard-mined set. To cancel the effects of the PA-ASPP module, we apply them to the DeepLabv3+ for direct performance comparison. Tab. 2 shows the quantitative performances. Compared with the original DeepLabv3+, the HEM can consistently improve the classification results. Note that the HEM is a general proposal. It can be used in varied settings for training segmentation networks.

3) EFFECTS OF THRESHOLDING REGIONS
For predicting patient-level results, we convert each whole-slide probability map into a feature vector with 5 threshold values, i.e., 0.5, 0.6, 0.7, 0.8, 0.9. Based on the regions, we train a lymph node classifier. To verify the effect, we show the visual results in Tab. 3. As we can see, at a low threshold, the final prediction can include noise for the   cancer localization. While at a high threshold, it will miss some key regions. Our multiple thresholds can effectively highlight the most of regions. Fig. 9 shows the effects with different thresholding values.

4) LIMITATIONS
We address the patient-level lymph node status classification problem by regional metastases segmentation of gigapixel whole-slide images using the divide and conquer strategy. The proposed approach shows superior performance when compared to its competitors. However, our method still shows low accuracy for ITC, and the model size is still very large for real applications. Besides, the patch extraction is based on the Otsu thresholding. In some cases of CAMELYON2017, the complex background is still be included.

V. CONCLUSION
In this work, we present a fully automatic metastases detection framework to predict pN-stage of patient-level lymph node statuses from the whole-slide histopathology images.
Several useful strategies are proposed to improve the regional segmentation accuracy and integrate multi-granularity representations for the final classification. Our method adopts a divide and conquer strategy and allows the model to process the whole-slide images. More specifically, our method first divides the whole-slide images into tumor-related patches, then performs tumor segmentation on regional patches, and integrates segmented patches to generate the segmentation of the entire WSI. More importantly, we propose a new deep segmentation network (DSNet) with pyramid attentions and scale-aware selection so that the expressive multi-level features can be emphasized to facilitate the feature refinement. Extensive comparative experiments with 5-fold crossvalidation on CAMELYON17 demonstrate that our proposed method can significantly outperform several state-of-theart methods in typical metrics. Due to the characteristics of digital pathological images, it's impossible to manually annotated large quantity qualified datasets for model training, [34] has proposed a weakly supervised deep learning framework to solve the issue. To enable clinical application, we will improve the patch sampling method to generate more valuable training examples; and we will improve the system framework to enable weakly supervised learning and semi-supervised learning. On the other side, according to the review article [35], the applications for DL based pathology analytics for Breast Cancer are still focused on tumor detection, the basic image analysis and at internal validation stage. To enable the advanced clinical applications, we will explore more advanced segmentation frameworks and backbone networks for the dedicated task, and we will build an end-to-end learning framework to predict the pN phase of the WSI and enable the accurate grading and subtyping clinical applications. JENQ-SHIOU LEU (Senior Member, IEEE) received the B.S. degree in mathematics and the M.S. degree in computer science and information engineering from the National Taiwan University, Taipei, Taiwan, and the Ph.D. degree on a part-time basis in computer science from the National Tsing Hua University, Hsinchu, Taiwan. He was with Rising Star Technology, Taiwan, as an Research and Development Engineer, from 1995 to 1997, and in the telecommunication industry (Mobitai Communications and Taiwan Mobile), from 1997 to 2007, as an Assistant Manager. In February 2007, he joined the Department of Electronic and Computer Engineering, National Taiwan University of Science and Technology (NTUST), as an Assistant Professor. From February 2011 to January 2014, he was an Associate Professor. Since February 2014, he has been a Professor. He served as the Department Chairman, from August 2017 to July 2020. His research interests include heterogeneous network integration, mobile service and platform design, cybersecurity, distributed computing, and green and orange technology integration. He has published extensively in these areas, with 78 SCI indexed journal articles, 65 conference papers or book chapters, and led 19 MOST projects, 28 industry-academia projects, and ten cross-university projects after he joined NTUST.