Researchers have developed a deep learning network architecture that achieves fully automated radiologist-level segmentation of breast tumours in MRI.
Segmentation
Radiologists undergo segmentation of breast tumours to obtain image features such as shape, morphology, texture and enhancement dynamics. Identifying these features can improve diagnosis and prognosis in breast cancer patients. Manual segmentation is labour intensive, yet to date, alternative reliable automated methods do not exist. The development of automated segmentation using deep network techniques has the potential to meet this clinical need.
Previously, researchers have applied deep learning methods to breast tumour segmentation and diagnosis in mammograms. However, unlike MRI, mammograms cannot determine the exact 3D location and volumetric extent of a lesion. Breast MRI has a higher diagnostic accuracy and outperforms mammograms in detecting residual tumours after neoadjuvant therapy. Currently, trained radiologists outperform automated methods on MRI segmentation.
A deep convolutional network
In this study, published as a preprint in arxiv, researchers hypothesised that by using a sufficiently large dataset to train a modern deep convolutional network, human-level performance could be achieved. The team retrospectively accessed 38,229 clinical MRI breast exams and portioned them into training and test data at random. The network was trained to distinguish between malignant and benign voxels, using 2,555 malignant breasts that were segmented in 2D by breast radiologists and 60,108 benign breasts as controls.
The team found that volumetric U-Net was the best-performing network on the training set. This network achieved radiologist-level performance on an independent test set of 250 breasts. It also had a Dice score of 0.77 (score of 1.0 corresponds to perfect overlap) and radiologist performance of 0.69-0.84.
Overall, a volumetric U-Net performs as well as expert radiologists at segmenting malignant breast lesions in MRI. Automated segmentation in the clinic has the potential to aid detection and diagnosis and improve overall workflow of radiologists.
Image credit: By shironosov – canva.com
