Breast-tissueMany patients undergoing breast-conserving surgery for the treatment of early-stage breast cancer require additional surgery because the surgeon was unable to locate and remove the entire tumour during the initial procedure. Improved tools to enable benign and malignant tissue to be differentiated may help solve this problem. Optical coherence tomography (OCT) is an imaging technique that can generate 3-D images to a depth of ~1 mm in tissue in real time. While benign adipose (fatty) tissue is easily identifiable in OCT images, benign dense tissue and malignant tissue are often harder to differentiate. To address this, functional extensions of OCT, such as attenuation imaging (which maps the rate of decay of light as it travels through tissue), have been developed and shown to improve contrast between benign and malignant dense tissues. The aim of this project is to use machine learning to combine OCT and attenuation imaging for automated breast tissue classification.
Principal investigatorKen Foo firstname.lastname@example.org
Area of scienceBiomedical Engineering
Applications usedPython (PyTorch)
20–30% of early-stage breast cancer patients undergoing breast-conserving surgery require additional surgery due to the suspected incomplete removal of the tumour during the initial procedure. These additional surgeries can have adverse effects on the physical and mental health of the patient, and also result in increased costs to the patient and healthcare system. To reduce the re-excision rate, tools that improve the surgeon’s ability to locate the tumour intraoperatively are required
OCT is an imaging technique that uses near-infrared light to rapidly construct high resolution (~10 μm) 3-D images to a depth of up to ~1 mm in tissue. Previous studies have shown that while OCT provides good contrast between dense tissue and fatty tissue in the breast, it can be more difficult to distinguish between malignant and benign dense tissue. To provide additional contrast, other OCT-based techniques such as attenuation imaging (which measures the rate at which light decays as it travels through tissue) have been investigated, with promising results. However, the automated classification of breast tissue by combining these techniques using machine learning has not yet been investigated.
The aim of this project is to classify breast tissue as benign or malignant based on the combination of OCT and attenuation imaging. Specifically, a convolutional neural network that outputs a classification based on the input of a multi-channel breast tissue image is being developed. Virtual graphical processing units (GPUs) available through the Nimbus cloud computing platform have enabled network training to be completed ~25x faster compared to a using a computer without a GPU. This has made it possible to tune the hyperparameters of the network more effectively, and thus maximise performance.
List of Publications
Multi-class classification of breast tissue using optical coherence tomography and attenuation imaging combined via deep learning (journal article submitted to Computerized Medical Imaging and Graphics).