Cover story for Pulse@UM Year 2021 Issue 1: Big Data & Artificial Intelligence in Medicine and Healthcare Research

Deciphering Radiomics Signatures of Breast Cancer – the BREOMICS approach

Education vector created by vectorjuice –

By Associate Professor Jeannie Wong Hsiu Ding, Professor Dr Kartini Rahmat, Professor Dr Ng Kwan Hoong & Associate Professor Dr Chan Chee Seng

According to GLOBOCAN 2018, 18.1 million new breast cancer cases and 9.6 million breast cancer deaths were reported worldwide. It is the second most common cancer, representing 11.6% of all cancers in females after lung cancer. Breast cancer, if detected and treated early, would result in a better prognosis for patients. Whilst mammography has been proven to reduce breast cancer mortality, its sensitivity decreases in younger women and women with mammographically dense breasts (Figure 1). 

Figure 1: Mammographic
images of a dense breast with
an ill-defined tumour
Figure 1: Mammographic images of a dense breast with an ill-defined tumour

Currently, histopathological diagnosis from a biopsy or surgical specimen is the gold standard for the determination of breast cancer subtypes. However, it has been associated with adverse clinical outcomes and variable responses between patients with the same diagnosis and treatment. Furthermore, biopsy is prone to under-sampling due to the heterogeneous nature of the lesion. 

In contrast, imaging modalities such as mammography, ultrasound, and magnetic resonance imaging (MRI) can visualise the whole breast lesion and its heterogeneous nature. Conventional radiological practise involves subjective perceptual skills in the diagnosis and stratification of breast cancer risks, leading to variations across interpreters and the adverse effect of fatigue.

In addressing subjective radiological interpreters’ limitations, computer-aided detection /diagnosis (CAD) systems based on different machine learning algorithms have been developed and deployed in some commercial mammography systems since 1998. In the last decade, deep learning-based artificial intelligence (AI) systems in medicine have also gained momentum.

Here, researchers from the Faculty of Medicine, in collaboration with the Faculty of Computer Science & Information Technology, took on the approach of exploring radiomics features of breast cancer. This project, code-named “BREOMICS” stands for (BREast radiOMICS), attempts to decipher breast cancer through the revolutionary use of radiology image datasets.

Radiomics is the extraction of a vast amount of quantitative features from digital medical images, not usually visible to the naked human eye, through high output computing to produce image-based tumour phenotyping. The central idea behind the concept of radiomics is that the quantitative image features (also called image phenotypes) reflect the tissue or disease. In other words, medical images are not just images; they are data that can be mined. 

Radiomics involves image acquisition -> image segmentation -> radiomics feature extractions -> feature reduction -> radiomics modelling. There are two different approaches to radiomics, the conventional handcraft radiomics and the deep learning-based approach (Figure 2). 

Figure 2: Deep learning-based radiomics
Figure 2: Deep learning-based radiomics

Handcraft or conventional radiomics mainly extract and select features such as histogram, morphology, wavelet feature from the medical images after the pre-processing steps such as denoising, normalisation and segmentation. Accurate segmentation is usually performed by radiologists. Hence, it is very human-dependent. Deep learning radiomics, on the other hand, employs end-to-end learning, a learning mechanism where the training of a system is represented by only a single model, bypassing the intermediate layers usually presented in conventional handcraft radiomics pipeline designs. However, the full advantage of deep learning radiomics needs to leverage on big data availability, the close inter-discipline collaboration between clinical experts, medical physicists and computer science specialists. 

Recent improvements in deep learning radiomics (DLR) extracting high-level features from medical imaging could promote the performance of computer-aided diagnosis (CAD) for cancer. Deep learning-based approach has been utilised in different parts of the radiomics process. In the era of precision medicine, BREOMICS aims for a total understanding of breast cancer which will uncover the intrinsic relationship between various radiologic imaging phenotypes with other breast cancer biomarkers, able to interrogate the whole breast lesion with a single “virtual biopsy”, providing an accurate differential diagnosis of breast lesions, towards a more precise and personalised treatment. 


1. Bray, F., et al. (2018). Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA: A Cancer Journal for Clinicians. 68(6): p. 394-424. 

2. Freer, P.E. (2015). Mammographic Breast Density: Impact on Breast Cancer Risk and Implications for Screening. 35(2): p. 302-315. 

3. Pesapane, F., et al. (2018). Artificial intelligence in medical imaging: threat or opportunity? Radiologists again at the forefront of innovation in medicine. European radiology experimental. 2(1): p. 35-35. 

4. Acharya, U.R., et al. (2018). Towards precision medicine: from quantitative imaging to radiomics. Journal of Zhejiang University. Science. B. 19(1): p. 6-24. 

5. Crivelli, P., et al. (2018). A New Challenge for Radiologists: Radiomics in Breast Cancer BioMed Research International. 2018: p. 10. 

6. Gillies, R.J., et al. (2016). Radiomics: Images Are More Than Pictures, They Are Data. 278(2): p. 563-577. 

7. Pang, T., et al. (2020). Deep learning radiomics in breast cancer with different modalities: Overview and future. Expert Systems with Applications. 158: p. 113501. 

Other posts: