BCRecommender System for Breast Cancer Diagnosis using Machine Learning Approaches

doi:10.21203/rs.3.rs-1985404/v1

Download PDF

Research Article

BCRecommender System for Breast Cancer Diagnosis using Machine Learning Approaches

https://doi.org/10.21203/rs.3.rs-1985404/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background: Early detection of breast cancer is challenging and necessitates more thorough observation. Researchers and medical professionals are always looking for ways to detect breast cancer early, systematically, and affordably. Other studies on cancer have developed a range of recommendation systems, such as those for food and medicine, disease prediction based on text and image data, and prognosis. The only drawback to such systems is that they are highly specialized and only accept specific types of data. To be more precise in diagnosis, we need systems that consider all aspects of the disease and provide continuous support at all the stages of breast cancer.

Methods: A recommendation model called "BCRecommender" has been put forth to diagnose and comprehend breast cancer using clinical and histopathological slides while taking feature-based classification into consideration. It has a layered architecture that accepts different reports (structured and unstructured data), and it produces results differently with respect to each layer. It is effective for cancer in all stages, from early to advanced.Existing breast cancer repositories are used to design a system.

Results & Conclusions:The model performance varies at different layers.In layer1 the bagging classifier achieves the highest accuracy of 61.06% while in layer2 bagging has the highest accuracy of 97.52%. Layer 3 has an accuracy of 97.39% after augmentation. During the test phase, confidence ranges from 60%-100%.In Layer 4 confidence varies from 50%-100%.

Recommendation System

Cancer

MRI Images

Cancer Diagnosis

CNN

Machine Learning

Drug Discovery

Prescriptive Analysis

Early identification and treatment of breast cancer is always valuable, however existing methods for early detection have moderate accuracy. In the majority of cases, cancer prognosis reaches a point where treatment is no longer possible. A substantial breakthrough in screening methods, early identification, and pharmacotherapy led to rise in survival rate.Doctors and scientists are attempting to repurpose existing datasets of cancer patients in order to uncover patterns for disease occurrence. This pattern analysis is carried out using well-known computational techniques known as Machine Learning (ML).

Artificial intelligence (AI) and ML laid the foundation for self-diagnosis tools by leveraging existing large data sets to address future issues of human disease identification at an early stage, particularly in cancer.[1]. These techniques can find patterns and trends between cancer occurrences in massive datasets, as well as make predictions related to cancer consequences. Customized medicine with specific recommendations is another ML application that has grown in significance for cancer prognosis[2].Recent years have also seen a rise in the use of AI-based systems for the early detection of breast cancer on various imaging modalities, including mammography, ultrasound, MRI, and biopsy histopathological images.Clinical tests like biopsies are the most common test for breast cancer after screening. In AI based clinical studies, classification models are developed using histological images of infected tissues removed during biopsy. This facilitates a quicker and more thorough understanding of inflammation as well as the kind and stage of cancer.

Breast Cancer Histopathological Image Classification (BreakHis) was proposed by [3] is made up of 9,109 microscopic examinations of breast tumor tissue with various magnifying factors.System tries to classify types of breast cancer. In a similar study,500 microscopy images of breast cancer were provided by [4] , 400 images were used as train dataset and the remaining 100 as test dataset. Results are divided into four groups by the classification model: normal, benign, in-situ, and invasive.

60 breast biopsy slides from breast cancer, ductal carcinoma in situ, atypical hyperplasia, and benign cases without atypia were included in the study's dataset, which was gathered between 2011 and 2014. The authors claimed to have discovered a good degree of pathologists' agreement (75.3 percent coherence) [5]. In a study [6] authors utilized 179 nonlinear microscopy (NLM) specimens to provide detailed assessment of cancer. The findings revealed that diagnoses based on NLM images were highly accurate.A remote mammary cancer examination device based on pathological images (92 images) was described by [7]. 4 ML algorithms were used to examine malignant & benign tissues.Based on images from fine needle aspirate (FNA) microscopy, [8] developed machine based cancer detection.K-means (KM), fuzzy c-means (FCM), Gaussian mixture model (GMM), and competitive neural network (CNN) were used in this study. The obtained accuracy varied from 96% to 100%. A study [9] introduced breast cancer histopathological annotation and diagnosis dataset BreCaHAD with 162 images.The ML-based system was developed to predict annotated into 6 categories ( mitosis, apoptosis, tumor nuclei, non-tumor nuclei, tubule, and non-tubule).

A hybrid recommendation system named Breast Cancer Recommender System (BCRS) [10] was developed to deliver all the information aspects related to cancer, through a single system. The hybrid system used a switching approach between the collaborative and content based techniques to handle the cold start issue. The articles are recommended on the basis of user profile similarity with the target user. If the target article has been already read by the neighboring users then their rating is predicted on the basis of ratings provided by these similar users. In case the target article has not been read by the neighboring users then the rating is predicted on the basis of the articles read by neighboring users from the same document category. The drawback was that the system could not be evaluated using real cancer patients.

The open source biological databases have encouraged the medical practitioners to study the diseases on the basis of genetic factors as opposed to the traditional symptoms based approaches.In line with this approach a gene based collaborative filtering algorithm [GeneCF] was proposed to assist the doctors to find the list of genes with respect to each patient for effective diagnosis[11]. A new evaluation metric named gene precision coverage was used to assess the effectiveness along with a cancer staining procedure of the genes from the Human Protein Atlas to verify the obtained results. AMP, SAA1, S100P, SPP1 and CY2A7 and AFP were the six genes which were listed out of which only S100P, SPP1 and CY2A7 and AFP were identified as cancer staining. In order to decide the number of genes to be recommended an experimental design was proposed which was further validated from RNA expression of both healthy and liver cancer patients .The results were further validated against the genes and samples from cancer staining of patients.

A Personal Health Information Recommender (PHIR) system was developed with two basic components a) PHIR search engine for searching health related data based on the patients’ medical profile and preferences b) Semantic annotator app for the experts/doctors/health professionals to add validated text/articles /multimedia content to the system[12] . The main advantage of the developed system was that even if two patients are similar in terms of ailment and treatment history then also they might get a different piece of information based on their individual preferences and contexts. Search engine feature with the addon recommendation ability to match the patients profile the relevant/validated information from the health experts marked the uniqueness of this system.

Recently many other research studies have developed recommender systems for generating personalized narratives in the form of audio/video[13],suggesting personalized treatments[14],prognosis [15] etc., but still there is room for developing recommender systems that can work in real-time and help in breast cancer prognosis.

We would explain a recommender model for breast cancer research specifically using a blended approach in this study. While designing the recommender model several data sources have been considered from different studies and arranged in a layered architecture to form layers with rules. Each model comprises different machine learning models and takes different kinds of input as per the condition required. This architecture tries to blend different conditions with data and provides layer by layer recommendations.

2.1 Methodology

This research has a four-stage methodology: Problem definition, Data collection, model selection for system design and result reporting.

Problem Definition and Objectives:

The foundation of current systems is clinical verifications using computational methods. The only issue with such systems is that they are extremely specialized and only accept a specific kind of data. All factors (such as cancer's symptoms, lifestyles, allergens, genetics, etc.) must be taken into account for a more accurate cancer diagnosis. Consequently, a blended approach in the recommendation system is required. This study's objective was to gather secondary data from various levels of cancer testing in order to combine that information with machine learning models to develop a recommendation model. We are working to incorporate symptoms and reports into this system; in the future, we plan to use secondary datasets to integrate lifestyles and genetics into this system.

(i) Data Collection & Preprocessing:

A review of literature revealed a wide range of breast cancer data. It combines clinical test results, pathology reports, genetics, and lifestyle information. Table 2 discusses the datasets used in this study and their sources.

The dataset from [16] was preprocessed using one hot encoding for categorical features while real valued features were scaled and normalized to obtain the final dataset. The dataset from [8] contains only real values for all features so scaling and normalization were performed to obtain the transformed dataset.

Table 1: Datasets utilized for study

Sources	Sample collection	Nature of data & Size	Model
[16]	Clinical report (biomarker of breast cancer)	• Numeric and categorical • Size 166 (both patients with and without cancer)	Ensemble -Data set is a combination of numeric & categorical data. In such cases, the ensemble has high performance and is quite robust .
[8]	fine needle aspirate(FNA)	-real values -Size 569
[3]	Histopathology images	-images -9,109 microscopic images of breast tumor magnifying factors :40X, 100X, 200X, and 400X Contains 2,480 benign and 5,429 malignant samples	CNN - Image data sets are processed using CNN which perform well with medical images
[9]	Histopathology images	162 breast cancer histopathology images	Mask R-CNN - This data set is masked with specific parts of infection .The problem here is image recognition and in such cases Mask CNN is suitable

(ii) Model Selection & Design

Model Selection: After data selection, best performing classification models were identified. Table 1 discusses data, selected model and reason of selection.

Layered architecture: In this study, a layered architecture with the blending of data has been proposed because, as we all know, breast cancer screening is an efficient mechanism for timely identification and to ensure a higher chance of a successful course of treatment. Table 2 discusses the layers, purpose & prerequisite, dataset & model, recommendation, and feedback in the proposed model.

Table 2: Layer Structure of Blended Recommendation System- BCRecomender

Layer No.	Name	Purpose & Prerequisite	Dataset & model	Recommendation	Feedback
1	Clinical Assessor	-Examines the presence of cancer -blood sample	Dataset: 166 Train/Test: 70:30 Model : Ensemble	-Possibility of cancer -If possibility is true, recommended for physical examination followed by FNA test	Outcomes with samples stored in the database after verification from experts.
2	FNA Test	-Examine the type of tumor -digitized image sample	Dataset: 569 Train/Test: 70:30 Model: Ensemble	-Possibility of lump or node to be cancerous -If possibility is true, recommended for examination of type of cancer and treatment	-Sample values with recommendations stored in the database after verification from experts.
3	Tumor Categorization	-Examine the type of tumor -digitized image sample	Actual Dataset: 2,013 Selected images:1390 of malignant types (consist of 200X images only) Train/Test: 70:30 Model: CNN	-Possibility of particular type of cancer - Annotate sample for better recognition	-Sample images with recommendation stored in database after verification from expert
4	Strain Recognition	-Recognize infected part among :Mitosis, Apoptosis, Tumor nuclei, Non-tumor nuclei, Tubule and Non-tubule	Dataset: 162 Train/Test: 162 with number of annotations: 23549 Model: Mask R-CNN	-Strain recognition in different parts	-Sample annotated images with recommendation stored in database after verification from expert

In the first layer, the dataset contains data from routine blood analyses - notably, glucose, insulin, HOMA, leptin, adiponectin, resisting, MCP-1, age and Body Mass Index (BMI) - which can all be utilized to foretell the presence of breast cancer. This is an entry layer which will try to screen cancer based on clinical data. The model used is an ensemble, trained on already existing data. The outcome of this model is likelihood of presence (1)/absence (0) of breast cancer. This model receives initial patient data during recommendation execution, and it produces binary results (presence or absence). We advance to the second layer and also record the case in the training dataset of layer one if the outcome is presence. The system also advises getting pathology testing because the model predicts a high risk of malignancy. No additional test results would be taken into account in the event of absence.

In layer 2, a dataset from the FNA test was considered to design the model. Features are extracted from the imageset that includes traits of the cell nuclei visible in the image. It is used to train the ensemble model of this layer. In the testing phase, the FNA of the new target (patient) is then predicted. After being obtained from the patient, FNA test data is transformed into numerical values.The outcome of this model will again confirm presence and absence of breast cancer. If target values reveal the presence, then target would be suggested to go for further test and shifted to layer 3.

CNN models are trained using histopathology images in layer 3. The dataset is taken from BreakHis. It contains breast tumor tissue of 4 types of benign and 4 types of malignant tumors. We used a malignant tumour dataset of 200X images (i.e., 1390 total images) in this system to train our model. 4 malignant tumors: carcinoma (DC), lobular carcinoma (LC), mucinous carcinoma (MC) and papillary carcinoma (PC).This decision was made in order to maintain system speed and direct attention toward cancer type detection. This model's output aids in classifying the type of cancer from images. The suggested cancer type is either PC, DC, LC, or MC. If the input image differs from one of these classifications, the system won't produce accurate results.

In layer 4, an image recognition model is created using breast cancer histopathology images named BreCaHAD (developed by the biomedical imaging community). A pathologist labeled or annotated the hematoxylin and eosin (H&E) stained histological images as mitosis, apoptosis, tumour nuclei, non-tumor nuclei, tubule, or non-tubule. For our system, we have adjusted the image annotations. Since BreCaHAD does not have annotated data from papillary carcinoma (PC) cases, this model has not been trained for PC. The rectangle boxes used to highlight the image portions in this model will provide a masked part.

Though the system is in its infancy, we will use the MYSQL database for importing the data from the dataset and thereby storing the generated recommendations. These recommendations can be used as input to the feedback module wherein the experts can verify the results.

iv) Results & Analysis: In layer1 the bagging classifier achieves the highest accuracy of 61.06% while in layer2 bagging has the highest accuracy of 97.52%. Layer 3 has accuracy 97.39% after augmentation.During test phase, confidence ranges from 60%-100%.In Layer 4 the confidence varies from 50%-100%.

Table 3: Performance of Each Layer in Recommendation System

Layers	Model performance
1	Ensemble classifier	Accuracy (mean)	Recall		Precision	F1-score
	Bagging (n_estimators=100, criterion="entropy", max_features="auto" )	61.06	73.0		73.0	74.0
	Adaboost (n_estimators=200,learning_rate=0.1 )	56.72	73.0		75.0	74.0
	Voting classifier	53.49	74.0		74.0	74.0
2	Ensemble classifier (n_estimators=100, criterion="entropy", max_features="auto" )	Accuracy (mean)	Recall		Precision	F1-score
	Bagging	97.52	94.0		94.0	94.0
	Adaboost (n_estimators=200,learning_rate=0.1)	95.60	95.0		95.0	95.0
	Voting classifier	55.56	96.0		96.0	96.0
3	CNN	Training Loss	Training Accuracy		Validation Loss	Validation Accuracy
	Model before augmentation	.4311	83.18		0.9860	66.91
	Model after augmentation	.0631	97.39		1.9344	65.11
4	Mask RCNN (resnet)	Training Loss	Evaluation Loss	Confidence
		0.045	5.541	50-100%

The BCRecomender system designed in this research is unique and tries to provide continuous assistance for faster cancer detection. As the system still is in its initial phase, there are certain constraints that need to be handled to make this system more flexible and extensive. This section discusses each layer with their advantages and limitations. Table 6 describes the detailed observations.

Table 4: Observations on BCRecommender

Layers	Utility	Constraints	Enhancement Needed
1	Support to analyze clinical data at initial phase of cancer	Training dataset is limited and imbalanced	Dataset can be expanded and balanced for better training
2	FNT test analysis based on existing data	Training dataset is imbalanced and numeric	• Balanced dataset • Image dataset can be combined for better prediction
3	Support in Recognizing cancer type	• Model was trained based on 4 types of cancer • 200X images were considered	• Images from other type of cancer can be combined to improve recommendations • 400X images can also be considered for evaluation.
4	Support strain detection	• Type of strains and images are limited	• Images from other type of cancer can be considered. • strain annotated images can be increased for better analysis.

The proposed BCRecommender is an effort to address the challenges related to early detection of breast cancer, thereby providing necessary medications and rehabilitation. One of the major challenges is the imbalance of the datasets which result in biased predictions. Secondly the deep learning model in the layered architecture of the recommender system requires a large amount of image data for training thereby affecting the accuracy of predictions. Most of the datasets either lack this in number or quality. As a part of future work we would consider other cancer datasets including gene expression or metabolomics data.

Davenport T, Kalakota R.( 2019). The potential for artificial intelligence in healthcare. Future healthcare journal. Jun;6(2):94..
Kourou K, Exarchos TP, Exarchos KP, Karamouzis MV, Fotiadis DI..(2015) Machine learning applications in cancer prognosis and prediction. Computational and structural biotechnology journal.1;13:8-17.
Spanhol FA, Oliveira LS, Petitjean C, Heutte L.(2015). A dataset for breast cancer histopathological image classification. Ieee transactions on biomedical engineering.;63(7):1455-62.
Aresta G, Araújo T, Kwok S, Chennamsetty SS, Safwan M, Alex V, Marami B, Prastawa M, Chan M, Donovan M, Fernandez G.(2019). Bach: Grand challenge on breast cancer histology images. Medical image analysis.;56:122-39.
Elmore JG, Longton GM, Carney PA, Geller BM, Onega T, Tosteson AN, Nelson HD, Pepe MS, Allison KH, Schnitt SJ, O’Malley FP.(2015).Diagnostic concordance among pathologists interpreting breast biopsy specimens. Jama.;313(11):1122-32.
Tao YK, Shen D, Sheikine Y, Ahsen OO, Wang HH, Schmolze DB, Johnson NB, Brooker JS, Cable AE, Connolly JL, Fujimoto JG.( 2014) .Assessment of breast pathologies using nonlinear microscopy. Proceedings of the National Academy of Sciences.;111(43):15304-9.
George YM, Zayed HH, Roushdy MI, Elbagoury BM. (2013). Remote computer-aided breast cancer detection and diagnosis system based on cytological images. IEEE Systems Journal.;8(3):949-64.
Kowal M, Filipczuk P, Obuchowicz A, Korbicz J, Monczak R.(2013).Computer-aided diagnosis of breast cancer based on fine needle biopsy microscopic images. Computers in biology and medicine.;43(10):1563-72.
Aksac A, Demetrick DJ, Ozyer T, Alhajj R.(2019) BreCaHAD: a dataset for breast cancer histopathological annotation and diagnosis. BMC research notes.;12(1):1-3.
Taban M. (2014).A recommender system for breast cancer patients (Doctoral dissertation, Memorial University of Newfoundland).
Hu J, Sharma S, Gao Z, Chang V.(2018) Gene-based collaborative filtering using recommender system. Computers & Electrical Engineering. 1;65:332-41.
Iatraki G, Kondylakis H, Koumakis L, Chatzimina M, Kazantzaki E, Marias K, Tsiknakis M.( 2018) Personal Health Information Recommender: implementing a tool for the empowerment of cancer patients. ecancermedicalscience.;12.
Ormel I, Onu CC, Magalhaes M, Tang T, Hughes JB, Law S.(2021). Using a Mobile App–Based Video Recommender System of Patient Narratives to Prepare Women for Breast Cancer Surgery: Development and Usability Study Informed by Qualitative Data. JMIR formative research.;5(6):e22970.
Katzman JL, Shaham U, Cloninger A, Bates J, Jiang T, Kluger Y. (2018) .DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network. BMC medical research methodology.;18(1):1-2.
Kanimozhi G, Shanmugavadivu P, Rani MM.(2020). Machine Learning‐Based Recommender System for Breast Cancer Prognosis. Recommender System with Machine Learning and Artificial Intelligence: Practical Tools and Applications in Medical, Agricultural and Other Industries.:121-140.
Patrício, M., Pereira, J., Crisóstomo, J., Matafome, P., Gomes, M., Seiça, R., & Caramelo, F. (2018). Using Resistin, glucose, age and BMI to predict the presence of breast cancer. BMC cancer, 18(1), 1-8

Download PDF

Version 1

posted

You are reading this latest preprint version

BCRecommender System for Breast Cancer Diagnosis using Machine Learning Approaches

Status:

Version 1

Abstract

Figures

Introduction

Bcrecommender

Observations And Discussion

Challenges and Future Work

References

Status:

Version 1