Histopathological Auxiliary System for Brain Tumour (HAS-Bt) based on Weakly Supervised Learning using a WHO CNS5-Style Pipeline

doi:10.21203/rs.3.rs-2575131/v1

Purpose

Classification and grading of central nervous system (CNS) tumours play a critical role in the clinic. When WHO CNS5 simplifies the histopathology diagnosis and places greater emphasis on molecular pathology, artificial intelligence (AI) has been widely used to meet the increased need for an automatic histopathology scheme that could liberate pathologists from laborious work. This study was to explore the diagnosis scope and practicality of AI.

Methods

A one-stop Histopathology Auxiliary System for Brain tumours (HAS-Bt) is introduced based on a pipeline-structured multiple instance learning (pMIL) framework developed with 1,385,163 patches from 1,038 hematoxylin and eosin (H&E) slides. The system provides a streamlined service including slide scanning, whole-slide image (WSI) analysis and information management. A logical algorithm is used when molecular profiles are available.

Results

The pMIL achieved an accuracy of 0.94 in a 9-type classification task on an independent dataset composed of 268 H&E slides. Three auxiliary functions are developed and a built-in decision tree with multiple molecular markers is used to automatically formed integrated diagnosis. The processing efficiency was 443.0 seconds per slide.

Conclusion

HAS-Bt shows outstanding performance and provides a novel aid for the integrated neuropathological diagnostic workflow of brain tumours using CNS 5 pipeline.

neuropathology

histology

deep learning

multiply instance learning

clinical practice

Tumours of the central nervous system (CNS) consume abundant healthcare resources worldwide[1–3]. The classification of CNS tumours plays a critical role in the clinic, as it is used to estimate patient prognosis and guide clinical care. This explains why numerous efforts in the field of CNS tumours have been made into correlational research to continuously move the discipline forward. Recently, the WHO issued the fifth version of the international standard for the classification of CNS tumours, which includes more molecular pathological parameters while correspondingly streamlining the histopathology section according to newly revealed findings[4–10]. As a result, the already stretched resources of neuropathologists will have to tilt toward molecular pathology and diagnostic information integration in the future. Such reform calls for an automatic histopathology scheme by which pathologists can be liberated from laborious clinical work.

Artificial intelligence (AI) technologies, particularly deep learning methods, have been widely used in the healthcare field in recent years to address the explosive growth of medical data, whose handling is far beyond humans’ capacity[11–13]. Additionally, the development and widespread use of digital slide scanners has provided an enabling platform for the application of modern technologies in the burgeoning pathology field[14]. Driven by these concepts[11, 13, 15], researches have continued to rapidly develop deep learning-based AI systems for use in pathology[16–19]. Accordingly, the convolutional neural network (CNN), a sophisticated deep learning method, has become the primary choice in the field to meet the demand of automatic neuropathology services in the CNS5 era. A foundational study[20] based on classical CNNs realized the binary classification of high-grade (HGG) and low-grade glioma (LGG). A later study[21] developed a ternary glioma grading model based on ResNet18 through transfer learning, but this approach suffered from inferior accuracy, especially when differentiating grade II and grade III gliomas. Our team previously developed a model based on the squeeze-and-excitation–based DenseNet with weighted cross-entropy (SD-Net_WCE) to classify five major histological subtypes of glioma; the model was trained on over 79,990 histopathology image patches from 267 patients and achieved an 87.5% patient-level accuracy on a testing dataset derived from 56 other patients with 17,262 image patches[22]. Related studies have gradually evolved from binary to quinary glioma classification, consistently demonstrating the great potential of AI in the pathological diagnosis of brain tumours. However, a common limitation in current research should be noted; namely, the scope of diagnosis is still confined to gliomas, which are indeed the most common type of adult brain tumour³. It is reasonable to begin with glioma in this manner, but that does not mean that the classification of a single type of disease should be the goal of AI-based pathology. However, given the limited computing power of existing devices and the advanced level of deep learning algorithms, it is unrealistic to include all CNS tumour entities at once. According to the classification criteria recommended by the WHO CNS5 standard, we established a nine-type histopathology brain tumour classification task as the first step toward a fully automated AI brain tumour pathological diagnosis model Details of classification could be found in Supplementary material 1.

Even this preliminary step in extending the diagnostic scope of AI to 9 types might carry several challenges. Covering more entities incurs a surge in the sizes of datasets, as well as the complexity⁸ and heterogeneity[23, 24] of specimens from brain tumours. That means weakly supervised annotation (especially multiple-instance learning MIL[25]), which was widely used to meet challenge of annotating labor caused by massive datasets[18], could not reconcile itself to highly complex multiple brain-tumour classification tasks. Moreover, merely upgrading the algorithm while neglecting practicality is similar to creating a castle in the sky. Building a truly practical artificial intelligence pathological diagnosis service product requires further consideration of ways to make the operation of the algorithm conform to clinical reality and integrate it into the routine workflow of pathological diagnosis. Finally, the system should not conflict with human interests, allowing it to be integrated into existing social and economic systems. In conclusion, a well-designed MIL algorithm framework is needed to realize the complex and heterogeneous nine-class histopathological diagnosis of brain tumours using AI. Additionally, it is necessary to covert the algorithm into a practical AI pathological diagnosis service product on the consideration of integration.

In this article, we hypothesized that a pipeline-structured MIL (pMIL) algorithm framework could address the impact of the complexity and heterogeneity of brain tumours on weak supervised learning through a "pipeline" way rather than an "end-to-end" manner by dividing the task into two subproblems: cell subtyping and grading. And pMIL could allow us to build a one-stop Histopathology Auxiliary System for Brain tumours (HAS-Bt) that includes slide scanning, image analysis and information management. We set overall accuracy as primary outcomes and supposed that pMIL could reach the level of 0.9. Since molecular pathology has been an indispensable component of the WHO CNS5[26], we also design a decision tree to form an integrated diagnosis automatically.

Dataset and partition

In this article, we structured the dataset from Huashan Hospital’s BioBank of Central Nervous System Diseases[27] (Ethics approval: KY2015-256) according to the following criteria:

Inclusion criteria: 1. comprehension of the research process, agreement to use the patient data, and signing of the informed consent form by the patient or his or her guardian; 2. tumour surgery (resection or biopsy) from 2013 to 2021; 3. postsurgical pathological diagnosis of glioma, ependymoma, metastasis or lymphoma.

Exclusion criteria: 1. unavailable slides (missing or broken) 2. Unqualified slide or insufficient tissue; and 3. a controversial diagnosis from an experienced neuropathologist.

In general, 1,038 slides were collected. A total of 770 slides served as the modeling data, while the remaining 268 slides were used for independent testing. Among the modeling data, 142 slides were used to train the feature extractors for cell subtyping and grading. The other 628 slides were used to form the feature aggregator responsible for final label inference. A 3-fold development procedure was applied to ensure the robustness of the procedure on different dataset partitions. Detailed dataset and partition were shown in Fig. 1.

Digitization Of Histopathological Slides

An industrial-scale whole-slide imaging scanner (CytoExplorer ZJ300-CS2, ZHONGJI BIOLOGICAL, Wuhan) was used to digitize the histopathological slides. The scanner is equipped with a high-precision electrical displacement platform for precise positioning of the slides. Image acquisition is performed with a high-resolution camera with a highly adaptable, three-dimensional focusing control technology, which was developed to achieve fast and accurate digital scanning of WSIs. Image stitching and fusion are carried out on the scanned patches from multiple views to form a complete WSI.

Standard Annotation Procedure And Committee

To fit the pMIL framework, we adopted an annotation software based on a box selection function module that can quickly label various components contained in a WSI by directly circling them and outlining key pathological features that support the hierarchical diagnosis.

The uneven understanding of pathological classification and familiarity with annotation software make it challenging to standardize the annotation quality among different annotators. We established a committee to ensure consistent, high-quality annotation. In total, three pathologists from Huashan Hospital formed the standard annotation committee in this study, with one neuropathologist with more than 15 years of experience and two junior neuropathologists. The two junior members read all the WSIs in the training set and perform the annotation procedure independently. If the opinions on one WSI disagreed with each other (less than 50% overlap in the cell subtyping annotation boxes or more than 50% discrepancy in the high-risk feature boxes), the leader was invited to review the disputed WSI and make a final decision.

Pipeline-structured Multiple Instance Learning (Pmil)

The pMIL framework can be summarized into two groups of units: (1) clinical task-driven feature extractors and (2) integration aimed feature aggregators. We further established a cell-subtyping network and a grading network based on these.

Feature extractor

The goal of feature extractors is to implement dimensionality reduction on complex original inputs by embedding high-dimensional inputs into a low-dimensional feature matrix. The tasks or training targets of the feature extractors are designed to identify specific clinical image markers for better interpretation of the decision-making procedure.

Feature aggregator

The purpose of the feature aggregator is to provide a permutation invariant attention-based decision given a bag of vectors through instance pooling and multihead attention.

Cell-subtyping network

The patches to be classified included five cell types, oligodendroglioma (O), astrocytoma (A), ependymoma (EPE), lymphoma (LYM), and metastasis (MET), and two auxiliary types, background (BG) and nontumoural tissue (NT). NT covers normal brain tissues, glioses, bleeding and tissue inflammation. First, a feature extractor reduces the dimensions, disassemble all patches and embeds them into a category feature map. Then, a feature aggregator with 8 ports is used to pool and sample the feature map of the category in question and finally provide a decision that satisfies permutation invariance through the multihead attention mechanism.

Grading network

MVP and NEC are typical imaging biomarkers for grading; the presence of such markers indicates a greater likelihood of malignant behavior. Labeling the location of MVP requires a highly labor-intensive, pixel-level annotation; thus, instead of labeling the specific locations, we cropped the WSI into several patches and identified whether such features were present in each patch, changing the segmentation task into a classification task. The classifier also embeds meaningful features for the patient-level decision process. A feature extractor will first generate a grading feature map. The subsequent feature aggregator pools not only the level feature map, but also the sub-typing feature map obtained in the previous stage, so that the multi-head attention mechanism can "think comprehensively" about the grading tasks.

Detailed contents and algorithms of pMIL framework for this section could be found in Supplementary material 2.

Alternative Displays Of Diagnostic Information

Exclusion of low diagnostic areas and cell-subtyping map generation

In pMIL, we could get a classification mask in the classification stage. The entities we can predict in this stage include the five tumour types (O, A, EPE, LYM, and MET) and the two auxiliary types (NT and BG). Naturally, we can superimpose the above classification mask on the original WSI to gain (1) a low diagnostic value area prompt (BG + NT) and (2) a cell topographic map (O + A + EPE + LYM + MET).

High-risk image marker identification

The processing idea we take provides more auxiliary functions through backtracking the attention mechanism. We further leveraged gradient-based class activation mapping (Grad-CAM) to locate these high-risk markers. Detailed contents are given in Supplementary materials 3.

Molecular Pathology: Integrated Diagnosis

The taxonomy of the WHO CNS5 innovatively improves the weighting of non-histopathological parameters, such as molecular pathology. Therefore, we reserved an input window (Fig. 2a) for users to provide molecular pathological parameters in the future. The system will automatically integrate the input diagnostic information through a built-in decision tree (Fig. 2b). Integrated diagnosis of all slides could be found in Supplementary material 4.

Evaluation Metrics

A confusion matrix is used to describe the performance of the analysis unit. Accuracy, sensitivity and specificity are used to evaluate the resolving power of each category, and the arithmetic means of the accuracy, sensitivity and specificity of each category are used to show the overall performance of the model. The confusion matrix is segmented to analyze each category[28] (Fig. 3c). The metrics are defined as follows:

Accuracy_type =\(\frac{{Nt}_{TP}+{Nt}_{TN}}{{Nt}_{TP}+{Nt}_{FP}+{Nt}_{TN}+{Nt}_{FN}}\)

Sensitivity_type =\(\frac{{Nt}_{TP}}{{Nt}_{TP}+{Nt}_{FN}}\)

Specificity_type =\(\frac{{Nt}_{TN}}{{Nt}_{TN}+{Nt}_{FP}}\)

Accuracy_overall =\(\frac{1}{T}\bullet {\sum }_{t=1}^{T}\frac{{Nt}_{TP}+{Nt}_{TN}}{{Nt}_{TP}+{Nt}_{FP}+{Nt}_{TN}+{Nt}_{FN}}\)

Sensitivity_overall =\(\frac{1}{T}\bullet {\sum }_{t=1}^{T}\frac{{Nt}_{TP}}{{Nt}_{TP}+{Nt}_{FN}}\)

Specificity_overall =\(\frac{1}{T}\bullet {\sum }_{t=1}^{T}\frac{{Nt}_{TN}}{{Nt}_{TN}+{Nt}_{FP}}\)

where \({Nt}_{TP},{Nt}_{FP},{Nt}_{TN},{Nt}_{FN}\)represent the number of true positives, false positives, true negatives, and false negatives of each type, respectively. T represents the total number of types; thus, T = 9 in this study.

Prototype Packaging And Time Period Test

Based on the WSI scanner used in the study, we optimized the using process and built a scanning module. The pMIL model and the auxiliary functions were packaged to form the kernel diagnosis module. The information management module included images storage, diagnostic information integration and display functions. These three components were packaged to form the HAS-Bt. We evaluated the time consuming for the operation of HAS-Bt through the independent testing set. The 268 slides were run on the system. Timer were started when slide was settled on the system. Totally three path points were recorded. A researcher who has not participated in prototype development was responsible for recording.

Classification performance of pMIL

To avoid overfitting the model to a specific part of the dataset in the training process, we used a 3-fold development process. In 3-fold cross-validation, the performance of the submodel of each fold with the corresponding validation set can be described as a sub confusion matrix. When these sub confusion matrices are stacked together, a synthesized confusion matrix is formed (Fig. 3a), approximatively describing the model performance on “a virtual validation set” composed of all 628 cases. The overall accuracy was 0.959 ± 0.019. The overall sensitivity was 0.814 ± 0.081. The overall specificity was 0.977 ± 0.012 (Table 1.1).

Table 1. Classification performance of pMIL

Table 1.1

Classification performance in three-fold cross validation
		Index
		Accuracy	Sensitivity	Specificity
Subtypes	O2	0.942675159	0.770114943	0.970425139
	O3	0.952229299	0.72	0.972318339
	A2	0.933121019	0.786407767	0.961904762
	A3	0.929936306	0.728395062	0.959780622
	A4	0.96656051	0.892857143	0.973776224
	MET	0.974522293	0.775862069	0.994736842
	LYM	0.98089172	0.945454545	0.984293194
	EPE	0.984076433	0.931818182	0.992592593
	NT	0.963375796	0.78	0.979238754
Average		0.958598726	0.814545523	0.976562941

The validation set is typically used to adjust the hyperparameters of the model in deep learning. However, the “virtual validation set” was an approximate estimation of the 628 actual slides. Therefore, the above evaluation of the model cannot completely reflect the classification performance. We re-evaluated the performance with an independent test set composed 268 unused slides (Fig. 3d) and calculated the performance metrics (Table 1.2). The overall accuracy was 0.938 ± 0.032, 2.17% lower than that obtained with 3-fold validation (Table 1.3). The overall sensitivity fell to 0.720 ± 0.071, for an attenuation of 11.57%. The overall specificity decreased to 0.965 ± 0.035, a decline of 1.20%.

Table 1.2

Classification performance in independent test-set
		Index
		Accuracy	Sensitivity	Specificity
Subtypes	O2	0.944029851	0.714285714	0.978540773
	O3	0.899253731	0.8125	0.911016949
	A2	0.865671642	0.6875	0.889830508
	A3	0.955223881	0.677419355	0.991561181
	A4	0.936567164	0.65625	0.974576271
	MET	0.96641791	0.870967742	0.978902954
	LYM	0.951492537	0.6875	0.987288136
	EPE	0.947761194	0.642857143	0.983333333
	NT	0.973880597	0.733333333	0.988142292
Average		0.937810945	0.720290365	0.964799155

Table 1.3

Comparison of three-fold cross validation and independent validation
		Index
		Accuracy	Sensitivity	Specificity
Dataset	3-fold	0.958598726	0.814545523	0.976562941
Dataset	Independent	0.937810945	0.720290365	0.964799155
Decline(%)		2.168559197	11.57150281	1.204611096

Alternative Display Of Diagnostic Information

Screening out low diagnostic areas

Reading and identifying cell type and malignancy requires a carefully zoomed-in examination. Our model achieved a sensitivity of 0.95 on BG patches and 0.92 on NT patches in the patch-level cell type analysis, potentially filtering out unwanted impurities and unrelated regions for pathologists and helping guide more attention to tumour-related regions during slide reading (Fig. 4b).

Cell-subtyping map

In the cell-subtyping network, an intermediate layer identifies the five types of tumours based on a patch-level analysis. Therefore, application of cell-type identification on WSIs could provide a regional proposal of representative subregions of specific CNS tumour types in the way that representative regions are predicted as certain cell types with higher confidence. (Fig. 4a)

High-grade feature location prompting

Another patch-level application of our system is the risk evaluation model, which could suggest histopathological high-risk imaging markers that can help differentiate HGG from LGG (Fig. 4c). The accuracy of the patch-level high-risk and low-risk evaluation reached 0.97 with 0.89 sensitivity.

The One-stop Histopathology Auxiliary System For Brain Tumour (Has-bt)

We encapsulated the WSI scanning module, the kernel diagnosis module and the information management module to form a prototype of the one-stop Histopathology Auxiliary System for Brain tumour (HAS-Bt). The system only needs to manually settle the slides on the machine, and the subsequent scanning and analysis process will be carried out automatically. We performed a time period test on the independent testing set. According to statistical analysis, our system needs 147.97s for focusing; 188.59s for scanning and 115.55s for analyzing. In summary, our system needs 443.0s to process one slide averagely. System appearance and time-consuming statistic is shown in Supplementary material 5.

The goal of this study was to develop an AI diagnosis model that could histopathologically classify nine types of brain tumour. We posed a key scientific question regarding alleviation of the labeling demands far beyond the scope of human observers when the size of the dataset balloons to cover more entities. In view of the great potential of MIL in the field of pathological image recognition, we designed a pMIL framework to disassemble the complex multiclassification problems of brain tumours into cell subtyping and grading subproblems. We collected 1038 slides covering nine types of brain tumour from the biological sample bank of Huashan Hospital and digitized the slides with a WSI scanner. Using this dataset, we successfully trained and developed model. On an independent testing set, the overall accuracy of the model reached 0.938, the overall sensitivity reached 0.720, and the overall specificity reached 0.965. Compared with the results of three-fold cross-validation, the pMIL model showed a 2.2% reduction in accuracy, a 11.6% reduction in sensitivity and a 1.2% reduction in specificity. We also developed three auxiliary functions: low diagnostic area screening (LDAS), cell subtype mapping (CStM) and high-grade feature activation mapping (HFAM). In addition, we designed a built-in decision tree to automatically integrate any molecular pathology parameters input by users. Finally, we assembled all the components above into a one-stop system and conducted a rudimentary trial to assess the time consumption of the prototype. The processing efficiency of the system was approximately 443.0 seconds per slide.

The outcomes of this paper agree with our expectation that a pipeline-structured framework could improve the performance of MIL models, which means that it is not greedy to rely on the convenience of weakly supervised learning for labeling, given the effects internal heterogeneity and complexity of brain tumours on deep learning. A pipeline structure is not the primary choice in the domain of deep learning because it is suspected of coupling computer vision with human intelligence. From a certain perspective, this practice stifles the potential of computer vision. However, as AI continues to make headways in the field of medicine, such a notion perhaps must be revised. Ethical principles ensure that "effectiveness", "practicality" and "safety" are always important in the field of medicine. The purpose of medical research is to improve medical care and to provide ever-increasing benefits to human health. Our study takes an experimental step to verify the rationality and effectiveness of introducing our new understanding of brain tumours into artificial intelligence to improve model performance. When more brain tumour entities are included in the future or relevant research in the field encounters similar limitations, a pipeline structure may be an alternative way of compromise given limited computer and data resources. Additionally, algorithms and models are means for improving the quality of diagnosis and treatment. Researchers should not be content with current levels of progress and development in algorithms. This paper focuses on thinking about the landing and application of artificial intelligence in an actual clinical environment. Therefore, we further developed several auxiliary functions and designed a decision tree to incorporate molecular pathology. Surveying problems in the clinical environment can make these products "useful" rather than just "informative". We simplified the workflow from the perspective of integration to guarantee the "convenience" of these products (See hyperlinked supplement 1- System demonstration for actual practice).

As mentioned before, MIL can be adapted to the recognition of histopathological images, so relevant research in the field is not new. Companella et al.[29] directly applied whole-slide labels to all patches on a dataset consisting of 44,732 slides and built a deep learning model by using the MIL strategy to binary classify and diagnose prostate tumours. Their study benefited from the large scale of the dataset, which could be assessed with weakly annotation and unsophisticated models. This method is difficult to use in cases involving limited datasets, and the fixed samplers in their model made it difficult to guarantee generalizability and interpretability. Lu et al.[30] designed a clustered limit attention multiply instance learning (CLAM) framework based on 884 pathological slides of renal cell carcinoma to obtain a ternary classification diagnosis. They used a mature DenseNet as the feature extractor and an aggregated attention mechanism for sampling and explored reproducing the MIL strategy through a more complex structure when the dataset is limited. However, their ternary classification and diagnosis was developed as an overlapping of two binary classifications, and thus they do not avoid the limitations of binary level complexity. In conclusion, the innovations of this article are as follows: (1) we explored the application of the MIL strategy for high-dimensional multiclassification (a nine tumour-type diagnosis); (2) we used a pipeline structure to perform highly complex tasks; (3) we adopted a moderately supervised annotation method to balance the labeling demands and data reversibility; and (4) we produced a one-stop system that fits real clinical work environments.

However, there are still several limitations to address. “The six guiding principles for the design and use of artificial intelligence (AI) in healthcare”, another noteworthy document recently issued by the WHO, notes the ethical and socioeconomic risks accompanying the benefits of AI on improving the delivery of healthcare and medicine. These observations precisely emphasize the pain points of the study. The first limitation is “insufficient”. The scope of the study is not sufficiently wide. A critical challenge in the next stages of development will be identifying how to perform deep learning for rare tumours with a limited dataset. Additionally, insufficient diagnostic information was included. The criteria for classification of WHO CNS5 includes much more elements besides molecular pathology, such as age and lesion location. Methods for combining additional modalities of clinical information related to diagnosis must be explored to create an AI brain tumour pathological diagnosis model that is closer to the reality of clinical work. The second limitation is “unsustainable”. We cannot guarantee that our system is sustainable and continuously effective. AI will inevitably face an extrapolation crisis caused by overfitting when applied in a novel environment, so an update mechanism is needed to provide sustainable and robust medical services for brain tumours. The third limitation is “verification”. The capability of the pMIL is evaluated on the test set, but this is equivalent only to a single-armed observational study; therefore, it is not surprising that the strength of evidence is far from sufficient to support the practical application of AI in the clinic.

Author contributions:

Conceived and designed the study: YM, LJ, TS, FS, DS and JW

Gathered the data, performed and analyzed the research: YM, LJ, TS, HC, SW, JL, YZ, JZ and FS

Wrote the paper: YM, LJ and TS

Revised and edited the manuscript: FS, DS and JW

All authors read and approved the manuscript.

Data and materials availability

All data, code, and materials used in the analyses are available. Please contact authors via email if necessary.

Competing interests:

Tianyang Sun, Feng Shi and Dinggang Shen are employees in United Imaging Intelligence. Yaping Zou and Jun Zhang are employees in Zhongji Biotechnology. These two companies have no role in designing and performing the surveillances and analyzing and interpreting the data. All other authors report no competing interest relevant to this article.

Funding:

This project was supported by Shanghai Municipal Science and Technology Commision “Science and Technology Innovation Action Plan” Project (No.22S31905400), Shanghai Shenkang Hospital Development Center (No. SHDC12018114).

Bray F, Ferlay J, Soerjomataram I, Siegel RL, Torre LA, Jemal A (2018): Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin, 68(6):394-424.
Chen W, Zheng R, Baade PD, Zhang S, Zeng H, Bray F, Jemal A, Yu XQ, He J (2016): Cancer statistics in China, 2015. CA Cancer J Clin, 66(2):115-132.
Ostrom QT, Cioffi G, Gittleman H, Patil N, Waite K, Kruchko C, Barnholtz-Sloan JS (2019): CBTRUS Statistical Report: Primary Brain and Other Central Nervous System Tumors Diagnosed in the United States in 2012–2016. Neuro-Oncology, 21(Supplement_5):v1-v100.
Louis DN, Wesseling P, Paulus W, Giannini C, Batchelor TT, Cairncross JG, Capper D, Figarella-Branger D, Lopes MB, Wick W et al (2018): cIMPACT-NOW update 1: Not Otherwise Specified (NOS) and Not Elsewhere Classified (NEC). Acta Neuropathol, 135(3):481-484.
Louis DN, Giannini C, Capper D, Paulus W, Figarella-Branger D, Lopes MB, Batchelor TT, Cairncross JG, van den Bent M, Wick W et al (2018): cIMPACT-NOW update 2: diagnostic clarifications for diffuse midline glioma, H3 K27M-mutant and diffuse astrocytoma/anaplastic astrocytoma, IDH-mutant. Acta Neuropathol, 135(4):639-642.
Brat DJ, Aldape K, Colman H, Holland EC, Louis DN, Jenkins RB, Kleinschmidt-DeMasters BK, Perry A, Reifenberger G, Stupp R et al (2018): cIMPACT-NOW update 3: recommended diagnostic criteria for "Diffuse astrocytic glioma, IDH-wildtype, with molecular features of glioblastoma, WHO grade IV". Acta Neuropathol, 136(5):805-810.
Ellison DW, Hawkins C, Jones DTW, Onar-Thomas A, Pfister SM, Reifenberger G, Louis DN (2019): cIMPACT-NOW update 4: diffuse gliomas characterized by MYB, MYBL1, or FGFR1 alterations or BRAF(V600E) mutation. Acta Neuropathol, 137(4):683-687.
Brat DJ, Aldape K, Colman H, Figrarella-Branger D, Fuller GN, Giannini C, Holland EC, Jenkins RB, Kleinschmidt-DeMasters B, Komori T et al (2020): cIMPACT-NOW update 5: recommended grading criteria and terminologies for IDH-mutant astrocytomas. Acta Neuropathol, 139(3):603-608.
Louis DN, Wesseling P, Aldape K, Brat DJ, Capper D, Cree IA, Eberhart C, Figarella-Branger D, Fouladi M, Fuller GN et al (2020): cIMPACT-NOW update 6: new entity and diagnostic principle recommendations of the cIMPACT-Utrecht meeting on future CNS tumor classification and grading. Brain Pathol, 30(4):844-856.
Ellison DW, Aldape KD, Capper D, Fouladi M, Gilbert MR, Gilbertson RJ, Hawkins C, Merchant TE, Pajtler K, Venneti S et al (2020): cIMPACT-NOW update 7: advancing the molecular classification of ependymal tumors. Brain Pathol, 30(5):863-866.
Topol EJ (2019): High-performance medicine: the convergence of human and artificial intelligence. Nature Medicine, 25(1):44-56.
Rong G, Mendez A, Bou Assi E, Zhao B, Sawan M (2020): Artificial Intelligence in Healthcare: Review and Prediction Case Studies. Engineering, 6(3):291-301.
Yu KH, Beam AL, Kohane IS: Artificial intelligence in healthcare. Nat Biomed Eng 2018, 2(10):719-731.
Niazi MKK, Parwani AV, Gurcan MN (2019): Digital pathology and artificial intelligence. Lancet Oncol, 20(5):e253-e261.
Acs B, Rantalainen M, Hartman J (2020): Artificial intelligence as the next step towards precision pathology. J Intern Med, 288(1):62-81.
Chen PC, Gadepalli K, MacDonald R, Liu Y, Kadowaki S, Nagpal K, Kohlberger T, Dean J, Corrado GS, Hipp JD et al (2019): An augmented reality microscope with real-time artificial intelligence integration for cancer diagnosis. Nat Med, 25(9):1453-1457.
Ehteshami Bejnordi B, Veta M, Johannes van Diest P, van Ginneken B, Karssemeijer N, Litjens G, van der Laak J, the CC, Hermsen M, Manson QF et al (2017): Diagnostic Assessment of Deep Learning Algorithms for Detection of Lymph Node Metastases in Women With Breast Cancer. JAMA, 318(22):2199-2210.
Coudray N, Moreira AL, Sakellaropoulos T, Feny D, Tsirigos A (2018): Classification and Mutation Prediction from Non-Small Cell Lung Cancer Histopathology Images using Deep Learning. Nat Med. 24(10):1559-1567. doi: 10.1038/s41591-018-0177-5.
Ertosun MG, Rubin DL (2015): Automated Grading of Gliomas using Deep Learning in Digital Pathology Images: A modular approach with ensemble of convolutional neural networks. AMIA Annu Symp Proc.2015:1899-908.
Yonekura A, Kawanaka H, Prasath VA-O, Aronow BJ, Takase H (2018): Automatic disease stage classification of glioblastoma multiforme histopathological images using deep convolutional neural network. Biomed Eng Lett. 8(3):321-327. doi: 10.1007/s13534-018-0077-0.
Truong AH, Sharmanska V, Limbӓck-Stanic C, Grech-Sollars M(2020): Optimization of deep learning methods for visualization of tumor heterogeneity and brain tumor grading through digital pathology. Neurooncol Adv. 2(1):vdaa110. doi: 10.1093/noajnl/vdaa110.
Jin L, Shi F, Chun Q, Chen H, Ma Y, Wu S, Hameed NUF, Mei C, Lu J, Zhang J et al (2021): Artificial intelligence neuropathologist for glioma classification using deep learning on hematoxylin and eosin stained slide images and molecular markers. Neuro-Oncology 23, 44-52.
van den Bent MJ (2010): Interobserver variation of the histopathological diagnosis in clinical trials on glioma: a clinician's perspective. Acta Neuropathol, 120(3):297-304.
O'Connor JP, Rose CJ, Waterton JC, Carano RA, Parker GJ, Jackson A (2015): Imaging intratumor heterogeneity: role in therapy response, resistance, and clinical outcome. Clin Cancer Res, 21(2):249-257.
Amores J (2013): Multiple instance classification: Review, taxonomy and comparative study. Artificial Intelligence, 201:81-105.
Okamoto Y, Di Patre PL, Burkhard C, Horstmann S, Jourde B, Fahey M, Schuler D, Probst-Hensch NM, Yasargil MG, Yonekawa Y et al (2004): Population-based study on incidence, survival rates, and genetic alterations of low-grade diffuse astrocytomas and oligodendrogliomas. Acta Neuropathol, 108(1):49-56.
Aibaidula A, Lu JF, Wu JS, Zou HJ, Chen H, Wang YQ, Qin ZY, Yao Y, Gong Y, Che XM et al (2015): Establishment and maintenance of a standardized glioma tissue bank: Huashan experience. Cell Tissue Bank, 16(2):271-281.
Sokolova M, Lapalme G (2009): A systematic analysis of performance measures for classification tasks. Information Processing & Management, 45(4):427-437.
Campanella G, Hanna MG, Geneslaw L, Miraflor A, Fuchs TJ (2019): Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature medicine, 25(8):1.
Lu MY, Williamson DFK, Chen TY, Chen RJ, Barbieri M, Mahmood F (2021): Data-efficient and weakly supervised computational pathology on whole-slide images. Nat Biomed Eng, 5(6):555-570.

No competing interests reported.

Histopathological Auxiliary System for Brain Tumour (HAS-Bt) based on Weakly Supervised Learning using a WHO CNS5-Style Pipeline

Status:

Journal Publication

Version 1

Abstract

Purpose

Methods

Results

Conclusion

Figures

Introduction

Material And Methods