MEAI: an artificial intelligence platform for predicting distant and lymph node metastases directly from primary breast cancer

Breast cancer patients typically have decent prognoses, with a 5-year survival rate of more than 90%, but when the disease metastases to lymph node or distant, the prognosis drastically declines. Therefore, it is essential for future treatment and patient survival to quickly and accurately identify tumor metastasis in patients. An artificial intelligence system was developed to recognize lymph node and distant tumor metastases on whole-slide images (WSIs) of primary breast cancer. In this study, a total of 832 WSIs from 520 patients without tumor metastases and 312 patients with breast cancer metastases (including lymph node, bone, lung, liver, and other) were gathered. Based on the WSIs were randomly divided into the training and testing cohorts, a brand-new artificial intelligence system called MEAI was built to identify lymph node and distant metastases in primary breast cancer. The final AI system attained an area under the receiver operating characteristic curve of 0.934 in a test set of 187 patients. In addition, the potential for AI system to increase the precision, consistency, and effectiveness of tumor metastasis detection in patients with breast cancer was highlighted by the AI’s achievement of an AUROC higher than the average of six board-certified pathologists (AUROC 0.811) in a retrospective pathologist evaluation. The proposed MEAI system can provide a non-invasive approach to assess the metastatic probability of patients with primary breast cancer.


Introduction
Breast cancer is one of the most common and deadly cancers in women. Although the average 5-year overall survival rate for breast cancer patients is over 90% and the 10-year survival rate is 83%, they generally have a favorable prognosis. Even so, once breast cancer metastasizes, the prognosis dramatically deteriorates (Tian et al. 2017). The 5-year survival rate is 99% for breast cancer patients without metastases, whereas this percentage falls to 85% in lymph node metastases and to only 25% in cases of distant metastases. Moreover, metastasis in breast cancer patients is not only an essential prognostic factor, but also an important factor influencing treatment decisions (Howlader et al. 2016). For the best medical care and a possibility at survival, it is essential to identify if there are lymph node or distant metastases in cases of breast cancer (Amin et al. 2017). Currently, sentinel lymph node biopsy is utilized medically as a standard method for axillary lymph node staging in clinically node-negative patients with lymph node metastases of breast cancer (Krag et al. 2007), but as an invasive procedure, it still has complications (Wilke et al. 2006). For distant metastasis of breast cancer, the current approach is based on imaging, but it has the disadvantage of not identifying early and ultra-early-stage tumors. Based on the non-invasive prediction, some recent studies have proposed molecular markers to explore the metastatic mechanisms in different cancers (Zhang et al. 2021b;Afkari et al. 2021;Huang et al. 2022b). However, these molecular markers have been validated only in some ongoing trials and have the serious drawback of being costly and time-consuming.
Recent advances in deep learning have facilitated the development of artificial intelligence systems. Artificial intelligence (AI) has had a surprisingly positive impact on the field of pathology, including breast pathology. Developing algorithms for learning patterns from labeled digital data using "deep learning"-based neural networks and feature engineering methods have provided promising results (Shen et al. 2021;Akkus et al. 2017;Lv et al. 2022;Jaber et al. 2020). Image analysis and more sophisticated AI-based tools have proven helpful in performing computational assessments such as lymph node metastasis identification, tissue segmentation for diagnosing breast cancer, prediction, and tumor-infiltrating lymphocytes (Greenwald et al. 2022;Huang et al. 2022a;Liu et al. 2021). The gold standard for disease diagnosis is histological pictures, which can reveal detailed information about tumor features, underlying molecular processes, and disease development. Deep learning techniques that integrate visual and subvisual information of recurrent patterns in complex images can address this issue, because human assessment relies primarily on visually visible features, making it difficult to fully utilize the complex and rich information in histological images (Feng et al. 2020;Bejnordi et al. 2017;Naik et al. 2020). For exploring breast cancer metastasis through image information, previous work has predicted the extent of regional metastasis of breast cancer on histopathological sections of sentinel lymph nodes (Cserni et al. 2004;Liu et al. 2019;Xu et al. 2021). In particular, histopathological sections have not been used to study regional and distant metastasis of breast cancer.
Based on the above research background, we believe that deep learning techniques analysis of histopathological sections can predict the presence of distant metastasis and lymph node metastasis in breast cancer patients. In this work, we present for the first time an artificial intelligence system ( Fig. 1) that can give predictive results for distant versus lymph node tumor metastasis in breast cancer patients. Additionally, we exclusively used binary breast-level cancer labels for training, which are automatically derived from pathology reports. The interpretability of our system enables clinicians to establish trust and gain a deeper understanding of its advantages and disadvantages (Fig. 2).
The proposed system is an improvement in several ways relative to previous work. First, to the extent that we know, our proposed system is the first AI platform to directly predict distant metastasis from tumors on pathological sections of primary breast cancer patients. Second, we conducted a pathologist evaluation comparing this AI system's diagnostic accuracy to that of six pathologists to grasp the potential value of this AI system in clinical practice. The area under the receiver operating characteristic curve (AUROC) and the area under precision-recall curve (AUPRC) were significantly higher for the AI system compared to the six pathologists. In addition, we used the AI platform to assist pathologists in mixed diagnostics. The results showed that the specificity of pathologist predictions was significantly higher with the assistance of the AI system, and the same level of sensitivity was maintained. Meanwhile, we explored the relative predictive ability of the AI system for distant metastasis and lymph node metastasis of breast cancer, respectively. The results showed that the AI system could achieve an ACC of 0.908 (95% CI 0.898, 0.918) for lymph node metastasis of primary breast cancer and an ACC of 0.869 (95% CI 0.853, 0.885) for distant metastasis. We subsequently confirmed that the AI system's performance was consistent across age groups and metastatic patients.

Datasets
This study is based on publicly available data from TCGA (The Cancer Genome Atlas) and breast cancer pathology data collected from the Affiliated Hospital of Jiangnan University. Experienced pathologists microscopically examined data. We chose primary tumor samples from these cohorts and filtered photos of H &E-stained, formalin-fixed, paraffin-embedded sections to choose images independent of tissue processing artifacts (air bubbles, section folding, and poor staining). Pathological data for which clinical information was not available and positive physical or imaging studies confirmed by pathology were also excluded. The patient recruitment workflow is shown in Fig. 1. We applied the OTSU thresholding method to localize tissue regions in each WSI and then extracted non-overlapping patches from the tissue regions of size 224 × 224 pixels at 20× magnification. Finally, approximately one million patches were extracted from 312 slides from all patients, each containing pathological tissue in at least half of the area. 64 of these 312 slides had metastases in the bone, 56 in the lung, 48 in the liver, 96 in the lymph, and the remaining 48 contained more than two metastases (Table 1). In addition, we obtained 520 pathological slides from patients with non-metastatic breast cancer and processed 1,885,624 slides in the same way. All these slides were taken from breast cancer patients' primary cancer site-the breast.

Data preprocessing
The color of digital tissue sections can vary depending on the source material, the staining procedure, the digital scanner, and fading due to long-term storage, which ideally can be eliminated. To examine WSIs utilizing a variety of sources as well as our framework, we first used SPCN (structure-preserving color normalization) (Vahadane et al. 2016) to normalize the staining of WSIs. We set the RGB channel cutoff for the backdrop color to [220,220,220], based on the statistical findings of the data, since the WSIs background that all scanner captured was not really white in the conventional sense. To determine the global stain color appearance matrix W, we sampled a number of distinct background patches with pixel ratios 0.2 in each WSIs proportional to the WSIs' original size. Also, in calculating the global stain density map matrix H, we chose 99.9% of strong pseudomajor values for each row vector. Our data normalization was performed at 20 magnification, effectively reducing the normalization time. Afterward, the same normalization was performed on the standardized TCGA data for WSIs from The Affiliated Hospital of Jiangnan University. Along with the aforementioned elements, additional clinical data was gathered, including the patient's age at diagnosis, estrogen receptor (ER) status, and progesterone receptor. These data were used to help pathologists determine whether breast cancer patients develop regional lymph nodes and distant metastases in conjunction with pathological images and thus to compare with our model performance.

The metastasis recognition convolutional neural network (MECNN) architecture
A large portion of existing work on weakly supervised WSI classification has been described as multi-instance learning (MIL) (Dietterich et al. 1997;Chen et al. 2013). In the MIL framework, slides (or WSIs) act as bags that constitute multiple instances that are hundreds or thousands of patches cropped from the slides. A slide is marked as positive if at least one instance is positive for the disease; otherwise, it is marked as negative.
Previous deep learning tasks based on pathology images have had large datasets for training to achieve optimal model performance (Fu et al. 2020;Diao et al. 2021). For our AI system to achieve similar results on small and mediumsized datasets, we designed an image augmentation block ( Fig. 3b), while employing the idea of multiple instances learning to generate several enhanced slices from the original pathology slice.The enhanced slices generated by the augmentation block form an augmentation bag, which is then fed into the feature extraction network together with the original bag.
Before inputting bags into the multi-instance learning network, each bag will be amplified. One of the augmentation blocks contains two parts, which utilize traditional image processing methods and deep learning image augmentation networks, respectively. For the first part of the augmentation block, we selected the image augmentation method Co-Mixup (Kim et al. 2021) in deep learning. Co-Mixup performs image blending on a batch of input images to generate a batch of enhanced images. For this purpose, Co-Mixup maximizes the significance by penalizing to ensure local Tumor detection and ME prediction in H &E histology. ME and non-ME, sections of patients with metastatic breast cancer versus non-metastatic sections. a A threshold segmentation network was trained as a pathological tissue area detector. Tumor regions were cut into square tiles (b), which were color normalized and sorted into ME and non-ME (c). d Another network is trained to classify ME and non-ME, called metastasis recognition convolutional neural network (MECNN). e This automatic pipeline was applied to held-out patient sets data smoothing. We use pre-trained Co-Mixup to randomly select a random number of slices in the same package for image augmentation, which in turn yields the augmentation bag. In the second part of the augmentation block, we uses traditional image processing methods (Orr et al. 2020), including random flip, random rotation ( −180 to 180), staining matrix perturbation ( −10 to 10), and random staining concentration (0.5-1.5), to achieve image diversification. We randomly select an indefinite number of slices in the original bag to obtain another augmentation bag by traditional image processing methods.
The bags (including augmentation bag and original bag) are utilized as the input of the feature extraction network after being retrieved from the augmentation block. In each bag, N feature vectors are retrieved using a convolutional neural network (CNN) model for N picture examples. The WSIs are augmented with augmentation blocks to form several augmented bags and feature vectors are obtained from the feature extraction network, which are used as the input to our designed multi-instance learning network (AB-MIL) (Ilse et al. 2018). The final calculation of breast cancer metastasis probability is: where Y a represents the probability of lymph node and distant metastasis in primary breast cancer patients. (Y 3 ) represents the class probability of the original bag output through the AB-MIL model, and the corresponding (Y 1 , Y 2 ) represent the augmentation bag class probability. Finally, the AI system obtained the results of patient classification based on the probability score calculated by MECNN.

Network feature extraction
During the MECNN design phase, we carefully investigated a number of classical deep learning backbone networks for the model's feature extraction section (Fig. 3c). We choose the classic VGG16 (Simonyan and ResNet50 also achieved the best ACC, sensitivity, specificity, PPV, and NPV in the independent test cohort when other metrics were considered. Finally, ResNet50 was selected as the backbone network for MECNN feature extraction, in which convolution, batch normalization (BatchNorm) and rectified linear unit (ReLU) constitute the basic modules of ResNet50, and ReLU is used as an activation function to give nonlinear capability. There is a maximum pooling layer between the essential modules to achieve downsampling.

Pathologist evaluation
To evaluate the effectiveness of the suggested AI platform with those of the pathologists, we conducted a set. Six board-certified pathologists with an average of 6 years of clinical experience were included in this investigation (Supplementary Table 3). Their decades of expertise ranged from 1 to 15. Image data and clinical information (including pathology reports, whether they received chemotherapy, etc.) were provided to the pathologists. For each pathology section in all examinations, pathologists were asked to use values 1, 2, 3, 4, and 5 for predictive scoring; a score of 0 was not allowed.

Mixed diagnostic models
We developed a mixed diagnostic model for each pathologist, whose predictions were produced by averaging those of the corresponding pathologists and the AI model, to investigate the potential advantages that the AI system might offer. The AI system outputted metastatic scores as Y a , and the pathologists' metastatic scores were used as their predictions Y p : Both Y a and Y p were normalized to zero mean and unit variance. In this research study, we set = 0.5 , and the eventual Y calculated by weighting was used as a mixed diagnosis between the AI system and the pathologist

Statistical analysis
In this research, we assessed how AI systems, pathologists and mixed models performed using the following assessment metrics: accuracy, area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve, sensitivity, specificity, negative predictive value (NPV) and positive predictive value (PPV). AUROC and AUPRC were used to assess the probability of AI system/ mixed diagnostic models' generation-predicted diagnostic accuracy and manual diagnostic scores. Manual diagnoses were rated 1-5: scores 1 and 2 were classified in the lowest suspicion category; scores 3, 4, and 5 were independently considered as those with increasing levels of suspicion. We used Python Scikit to calculate AUROC and AUPRC.
In addition, we evaluated the binary predictions of AI systems, mixed models, and manual diagnoses using sensitivity, specificity, NPV, and PPV. These metrics are commonly used to assess the diagnostic accuracy in clinical studies. In this study, for each breast pathology section, the AI system and mixed model generated probability scores representing the likelihood of the presence of primary breast cancer metastasis. We dichotomized these scores to generate binary predictions by selecting a score threshold to distinguish between positive and negative decisions. To calculate sensitivity, we dichotomized the probabilistic predictions of the AI system to match the specificity of the average manual diagnosis. To calculate specificity, PPV, and NPV, we dichotomized the AI system's probabilistic predictions by matching the average pathologist's sensitivity. We similarly dichotomized the predictions of each mixed diagnostic model using the sensitivity/specificity of the respective manual diagnoses. For all evaluation metrics, we estimated confidence intervals of 95% by 1000 bootstrap method iterations. The p values were computed using a one-tailed permutation test with a statistical significance threshold set at 0.001. The calculation methods for all indicators are described in the Supplementary formula description of the Supplementary Material.

Study population characteristics
The total patients were divided into three subsets, the training set, the test set, and the mixed set. There were 437 patients in the training cohort, 187 in the independent test cohort, and 208 patients in the mixed set. The mean ages of patients in the training, test, and mixed sets were 57.6 (26-90 years), 56.7 (22-87 years), and 54.2 (28-82 years) years, respectively. There was no significant distinction in the significant characteristics between cohorts of patients with and without metastatic breast cancer (Table 2). 104 WSIs (12.5%) of all pathological sections collected were from patients with triple-negative breast cancer, and the remaining 728 (87.5%) were from patients with general breast cancer. Among these sections, 52 triple-negative breast cancer pathology sections were included in the training set with a weight of 50.0% and 28 in the test set with a weight of 26.9%. 26 (23.1%) were included in the mixed set. The independent test cohort, mixed cohort, and detailed characteristics of these cohorts did not correlate with one another, as indicated in Table 2.

MEAI-assisted ME-state assessment workflow
The AI platform workflow interface is shown in Fig. 1. After importing the WSIs, network1 is triggered to identify the tissue regions outlined in the slice and screen out the blank and impurity regions. The tissue areas are then cut into square tiles. The collected tiles are color normalized by the AI method and then manually screened by the pathologist, after which standardized sections are obtained. Subsequently, the tiles are fed into a neural network2 (MECNN) to obtain the metastatic status of breast cancer patients. The trained AI system outperformed the specificity and positive predictive value (PPV) of six attending pathologistsm while keeping the same sensitivity and negative predictive value level when we compared it to experienced pathologists for prediction (NPV).

AI system performance
The main training set of the experiment included 437 WSIs collected by TCGA and the Affiliated Hospital of Jiangnan University. Each image was cut into 224 × 224-pixel sizes and normalized. The MECNN architecture's metastatic recognition model was subsequently trained end to end under supervision using the AI system. Thereafter, a preliminary test set off an extra 187 slice images was used to put the AI system to the test (metastases 75; no-metastases 112). Ultimately, in the test set, the AI system achieved an AUROC of 0.934 (95% CI 0.922, 0.945) in identifying metastases in the presence of malignant lesions (Fig. 4). In particular, we stratified patients by metastatic site, age, and biomarker characteristics and evaluated the performance of the AI model in these subgroups (Supplementary Table 6). Our AI system recognizes lymph node metastases of primary breast cancer and achieves advanced performance in identifying ACC for distant metastases (Supplementary Table 7). The AI system maintained high diagnostic accuracy in all subgroups. Meanwhile, to investigate the effect of the augmentation block in MECNN (Fig. 3) on the predictive ability of the AI system, we conducted ablation experiments on the test set. The results showed that the AI system (AUROC 0.934, 95% CI 0.922, 0.945) was more accurate than the AI system without augmentation block (AUROC 0.897, 95% CI 0.880,0.914) and differed in prediction accuracy (Supplementary Table 2). We also explored the effect of the size of training slices on the performance of the AI systems. We observed that 224 × 224 versus 1024 × 1024 slice size resulted in better AUROC (Supplementary Table 1). However, larger images in the virtual training network consumed geometrically more computational resources during training, at 224 224 (AUROC 0.934, 95% CI 0.922, 0.945). For the sake of model generalization, we choose a smaller slice size

Model comparison
We measured the predictive power of the AI system on a test set and achieved the best results (AUC 0.934) for our model compared to several different traditional models. For data processing, we first identified the region containing the pathological tissue and subsequently cropped this region into small slices of 224 × 224 pixels. Around 200,000 small slices were finally selected to train the naive Bayes model, support vector machine (SVM), decision tree model and K-nearest neighbors (KNN) classifier. The trained naive Bayes models are naive Bayes using Bernoulli prior. SVM models employ grid search to determine the optimal kernel size and soft margins, and nonlinear radial basis functions are used as the kernel functions. KD tree search is used by KNN classifiers to determine the Euclidean distance to nearest neighbors. The decision tree model is a classification and regression tree (CART). The above steps were repeated ten times to prevent the randomness of individual experiments.
The results show that the SVM model achieves the best predictive power among the traditional machine learning algorithms (Fig. 5). However, compared to the AI system we developed, the evaluation metrics (ACC, AUC, sensitivity, specificity, PPV and NPV) were different (Fig. 6).

Pathologist evaluation
To compare the performance of the AI system with that of the pathologist, we also tested the pathologist's evaluation ability on the test set. Pathologists were informed that the study data set was enriched for metastatic cancer, but not the degree of enrichment. Six board-certified pathologists scored each breast based on breast pathology reports, sections, and clinical information from the corresponding patient. In Supplementary Table 3, the experience of the pathologists is presented. The patient's age, pathology reports that showed suspicious findings, and clinical diagnostic reports that identified any palpable regions of worry or pain were among the background details that pathologists were supplied. In contrast, no background information was given to the AI system. By comparing each pathologist's judged results to the ground truth data, we were able to determine receiver operating characteristic (ROC) curves and precision-recall curve for each pathologist (see "Statistic analysis" method). These six pathologists had an average AUROC of 0.811 and an average AUPRC of 0.789. The AI system in this investigation outperformed the performance of general pathologists, achieving a higher AUROC of 0.934 (Fig. 7). We similarly contrasted the specificity and sensitivity attained by the pathologists and the AI system. The six pathologists' mean specificity was 0.704 (95% CI 0.674, 0.734), and their mean sensitivity was 0.770 (95% CI 0.750, 0.790). In comparison to the pathologist's mean specificity, the AI system had a higher average specificity of 0.879 (95% CI 0.866, 0.891), with an incremental rise in specificity of 0.109 (95% CI 0.079, 0.139) (Supplementary Table 4).

Prospective medical clinical applications
On the basis of building a mixed diagnosis model of AI system and pathologists, we set up a mixed set (Table 2) to evaluate the potential of AI platform in strengthening pathologists' diagnosis. Each mixed diagnostic model's predictions were generated as the equally weighted average of the AI system and the manual scoring criteria (see "Methods" section "Mixed diagnostic models"). This investigation demonstrated that adding the forecasts of the AI system improved the performance of all predictions (Supplementary Table 5 (Fig. 8).
In conclusion, we found that the AI platform has a very high potential to assist inexperienced pathologists in the diagnosis of lymph node versus distant metastasis of breast cancer. Inexperienced pathologists have superior diagnostic results with the assistance of our AI system. As can be seen in Supplementary Table 5, in pathologists 1, 2, and 3 with 1, 3, and 4 years of study, respectively, their prediction accuracy differed from the AI system by about 15% without the assistance of the AI system. This gap decreased by 1/3 when AI was added to the mixed evaluation. These results demonstrate the potential of our AI platform for automated detection of lymph node and distant cancer metastasis screening in patients with primary breast cancer.

Discussion
In this work, we suggest an AI system that is at least as good as pathologists with 15 years of experience, capable of automatically detecting the presence of lymph node metastases and distant metastases (including bone, lung, liver, and metastases to more than two organs) in patients with primary breast cancer. The AI system was trained and assessed using a whole-slide images dataset gathered from the Affiliated Hospital of Jiangnan University and the TCGA public database, and it maintained a high level of diagnostic accuracy across cases. We confirmed its applicability in patient populations with different demographic compositions and institutions by validating its performance on a delineated test set. We conducted several advanced studies. First, in a pathologist evaluation, we found that the AI system outperformed pathologists with extensive experience. The six pathologists' average sensitivity was 0.770 (95% CI 0.750, 0.790), and the average specificity was 0.704 (95% CI 0.674, 0.734). Pathologists' sensitivity in our investigation was in line with their sensitivity in actual clinical practice. In the evaluation of pathologists, the AI system was able to detect distantly and lymph node metastases in breast cancer patients with superior sensitivity (0.909, 95% CI 0.872, 0.917) compared to pathologists while obtaining higher specificity (0.879, 95% CI 0.866, 0.891) and higher PPV (0.907, 95% CI 0.895, 0.918). The AI system outperformed all six pathologists in terms of AUROC (0.934, 95% CI 0.922, 0.945) and ACC (0.898, 95% CI 0.885, 0.910).

Fig. 5
The results of the traditional approach employing a confusion matrix on the testing set ( n = 187 WSIs). a Naive Bayes, b decision tree, c KNN, d SVM Fig. 6 Comparison of SVM, decision tree (DT), KNN, naive Bayes (NB) with AI system performance metrics in 10 repetitions of validation (in each repetition, we randomly select 75% of the data as training data and the remaining 25% as test data. Each The average of each indicator was calculated from 10 replicates to avoid the effect of data bias). a Accuracy, b AUC, c sensitivity, d specificity, e PPV and f NPV Fig. 7 The performance of the AI system on the mixed set ( n = 187 WSIs) using the ROC curve (a) and precision-recall curve (b). The AI achieved 0.934 (95% CI 0.919, 0.956) AUROC and 0.929 (95% CI 0.907, 0.951) AUPRC. Each red triangle represents a single pathologist, and the green triangles correspond to the average pathologist's performance 1 3 The possibility of using artificial intelligence systems to aid pathologists in analyzing metastatic breast cancer is another highlight of this work. We put forth and assessed a mixed diagnostic model that combines pathologists' predictions and computer algorithms. The study showed that this adjunctive approach improved the diagnostic accuracy of all six pathologists. In this present investigation, we demonstrated that the mixed diagnostic model raised the sensitivity of pathologists to 0.806 (95% CI 0.790, 0.821), with a relative increase of 0.036 (95% CI 0.020, 0.051). The mixed diagnostic model also improved the mean PPV of pathologists to 0.831 (95% CI 0.814, 0.848), an augmentation of 0.046 (95% CI 0.029, 0.063). These results suggest that our AI system can assist pathologists in making metastatic determinations of breast cancer and reduce missed diagnoses.
Eventually, we made a technical contribution to deep learning methods for medical image analysis. While previous deep learning methods have extracted features from pathological sections at the metastatic location for prediction Fan et al. 2019), our work extracts sections directly at the primary site of breast cancer patients for metastatic tumor prediction. In addition, most similar studies on predicting cancer metastasis, cancer subtype classification, and biological pathway activity are trained on large datasets, which not only require considerable computational resources, but also take days or Fig. 8 Pathologists' abilities and mixed diagnostic models' performance. We reported the mean values and 95% confidence intervals of ACC (a), AUROC (b), sensitivity (c), specificity (d), PPV (e), and NPV (f) of six pathologists and the mixed diagnostic models on the mixed set ( n = 208 WSIs). Each mixed diagnostic model's predictions are weighted averages of the evaluation ratings of each pathologist and the probabilistic forecasts of the AI (see "Methods" section, "Mixed diagnostic models") even months (Kather et al. 2019;Qu et al. 2021;Lin and Jeng 2020). Therefore, there is a need to develop a reliable AI system for clinical use that does not require training and validation on large-scale datasets while ensuring that the network can function appropriately under the various scenarios encountered in clinical practice. In this study, our proposed approach to augmenting the network allows our deep learning network to achieve similar performance on small datasets that can be used on extensive data.
Despite helping to improve the diagnosis of metastatic breast cancer, our study had certain drawbacks. While our AI system can perform similarly on short and large datasets, and can be taught on both, a potential research avenue is to evaluate it using large multi-class datasets also. Moreover, all the data collected is that each lesion finding is associated with only one image. In contrast, in clinical practice, technicians often acquire multiple slices of suspected malignancy from different angles and in different locations (Guo et al. 2019;Kundu et al. 2020;Zhang et al. 2021a). There will probably be some performance variance in AI between various test sets as a result of this heterogeneity in image acquisition. Future studies could focus on the pathological imaging of the same patient at different times and be used to assess suspicious morphological changes.
Regardless of these shortcomings, we believe this study presents the first AI for the prediction of distant metastasis and lymph node metastasis in breast cancer as a meaningful contribution. These findings give us confidence that our AI system can deliver highly accurate predictive performance and significantly improve pathologists' ability to diagnose diseases. The next step of our research will be to prospectively validate our AI system before practical application, testing it on a large number of multi-center, multi-type patient data. Given the large number of breast cancer patients referred for screening worldwide each year, the potential impact of our system in this area is enormous. Ultimately, the system aims to help breast cancer patients detect the possibility of tumor metastasis in advance and improve their survival and quality of life.