Machine Learning Methods for Breast Cancer Diagnostic

Shahnorbanun Sahran, Ashwaq Qasem, Khairuddin Omar, Dheeb Albashih, Afzan Adam, Siti Norul Huda Sheikh Abdullah, Azizi Abdullah, Rizuana Iqbal Hussain, Fuad Ismail, Norlia Abdullah, Suria Hayati Md Pauzi, Nurdashima Abd ShukorIntechOpen 2018 · DOI · OpenAlex

Abstract

This chapter discusses radio-pathological correlation with recent imaging advances such as machine learning (ML) with the use of technical methods such as mammography and histopathology. Although criteria for diagnostic categories for radiology and pathology are well established, manual detection and grading, respectively, are tedious and subjective processes and thus suffer from inter-observer and intra-observer variations. Two most popular techniques that use ML, computer aided detection (CADe) and computer aided diagnosis (CADx), are presented. CADe is a rejection model based on SVM algorithm which is used to reduce the False Positive (FP) of the output of the Chan-Vese segmentation algorithm that was initialized by the marker controller watershed (MCWS) algorithm. CADx method applies the ensemble framework, consisting of four-base SVM (RBF) classifiers, where each base classifier is a specialist and is trained to use the selected features of a particular tissue component. In general, both proposed methods offer alternative decision-making ability and are able to assist the medical expert in giving second opinion on more precise nodule detection. Hence, it reduces FP rate that causes over segmentation and improves the performance for detection and diagnosis of the breast cancer and is able to create a platform that integrates diagnostic reporting system.

Introduction

Breast cancer remains a leading cause of mortality among women globally. While early detection through screening is the most effective way to improve survival rates, the current diagnostic pipeline faces significant challenges. Radiologists often deal with a "needle in a haystack" scenario; in a typical screening of 1,000 mammograms, only a handful are cancerous. This high volume leads to fatigue and a miss rate of 10–30%. Conversely, the high rate of false positives—where the Positive Predictive Value is often below 35%—results in a surge of unnecessary and expensive biopsies.

Furthermore, the subsequent histopathological analysis is hindered by subjectivity. Pathologists must manually grade tissue samples, a process prone to inter-observer and intra-observer variation. This paper addresses these inefficiencies by proposing two machine learning-based systems: Computer-Aided Detection (CADe) for radiology and Computer-Aided Diagnosis (CADx) for pathology.

Methodology: CADe and CADx Systems

The researchers developed a dual-layered approach using Support Vector Machines (SVM) to provide an objective "second opinion" at different stages of the diagnostic process.

Overview of the CADe and CADx diagnostic framework.

The SVM Rejection Model (CADe)

The primary goal of the CADe system is to reduce false positives in mammogram screenings. Traditional segmentation methods, such as the Chan-Vese algorithm initialized by Marker Controlled Watershed (MCWS), often "over-segment," incorrectly flagging benign tissue as suspicious. To solve this, the authors introduced an SVM Rejection Model. This model acts as a post-segmentation filter, analyzing small "sample windows" (ranging from 7x7 to 13x13 pixels) to distinguish between actual nodules and noise.

Detailed workflow of the SVM Rejection Model training and testing phases.

Ensemble Framework for Histopathology (CADx)

For the diagnostic phase, the researchers moved away from a single "black box" classifier in favor of an Ensemble Framework. This system mimics the expert eye of a pathologist by breaking down tissue images into four semantic biological components: lumen, nuclei, cytoplasm, and stroma. Four separate SVM "specialists" are trained, each focusing on the features of a specific component. Their individual predictions are then aggregated using a "product rule" to determine the final diagnostic grade (Low vs. High grade).

Data and Results

The study utilized datasets from the UKM Medical Centre (UKMMC):

  • Mammography Dataset: 170 images from 109 patients were used to evaluate the CADe rejection model.
  • Histopathology Dataset: 110 high-resolution images (4140 × 3096 pixels) at 40× magnification, derived from 32 breast carcinoma cases. These were annotated by three expert pathologists using the Bloom–Richardson grading system.

Key Insights

  • False Positive Reduction: The CADe rejection model demonstrated significant efficacy. By using a 13x13 window size, the False Positive (FP) rate was reduced from 0.196 to 0.040. The researchers noted that larger window sizes provided better context, leading to higher specificity.
  • Diagnostic Accuracy: The CADx ensemble achieved a classification accuracy of 90.8% and an Area Under the Curve (AUC) of 90.7%.
  • Ensemble Superiority: The ensemble approach outperformed general CAD methods by approximately 15% in AUC. This suggests that training models on specific biological structures captures the "semantic meaning" of tissue architecture more effectively than abstract pixel patterns.

Discussion and Conclusion

The proposed machine learning methods offer a robust platform for improving breast cancer detection and diagnosis. By focusing on "rejection" to clean up radiological noise and "specialization" to automate pathological grading, these tools can significantly reduce human error and clinical fatigue.

While the results are promising, the authors note some limitations. The datasets for mammography and pathology were not linked to the same patients due to institutional constraints, and combining features from all tissue components can lead to high feature density and potential overfitting. Future work aims to integrate these data streams into a unified diagnostic reporting system. Ultimately, these systems are intended to assist, not replace, specialists, allowing them to focus on complex cases while AI handles routine screenings.

Summary updated

Abstract

This chapter discusses radio-pathological correlation with recent imaging advances such as machine learning (ML) with the use of technical methods such as mammography and histopathology. Although criteria for diagnostic categories for radiology and pathology are well established, manual detection and grading, respectively, are tedious and subjective processes and thus suffer from inter-observer and intra-observer variations. Two most popular techniques that use ML, computer aided detection (CADe) and computer aided diagnosis (CADx), are presented. CADe is a rejection model based on SVM algorithm which is used to reduce the False Positive (FP) of the output of the Chan-Vese segmentation algorithm that was initialized by the marker controller watershed (MCWS) algorithm. CADx method applies the ensemble framework, consisting of four-base SVM (RBF) classifiers, where each base classifier is a specialist and is trained to use the selected features of a particular tissue component. In general, both proposed methods offer alternative decision-making ability and are able to assist the medical expert in giving second opinion on more precise nodule detection. Hence, it reduces FP rate that causes over segmentation and improves the performance for detection and diagnosis of the breast cancer and is able to create a platform that integrates diagnostic reporting system.
BibTeX
@inproceedings{sahran2018machine,
  author = {Shahnorbanun Sahran and Ashwaq Qasem and Khairuddin Omar and Dheeb Albashih and Afzan Adam and Siti Norul Huda Sheikh Abdullah and Azizi Abdullah and Rizuana Iqbal Hussain and Fuad Ismail and Norlia Abdullah and Suria Hayati Md Pauzi and Nurdashima Abd Shukor},
  title = {{Machine Learning Methods for Breast Cancer Diagnostic}},
  booktitle = {IntechOpen},
  year = {2018},
  doi = {10.5772/intechopen.79446},
}
Figure