Monday, October 13, 2014

Machine Learning for Semi-Automatic Segmentation of Breast Lesions from MRI

I have recently had another journal article published on machine learning. This one is focused on a machine learning formulation I developed as a simplification of a technique I published earlier on - you can access information on the earlier publication here. The technique I have just had published (accessed from here) is a machine learning based tool for radiologists allowing them to delineate (or segment) a suspicious lesion from a breast MRI examination with a few seconds of user interaction to define tissue they are interested in analyzing further and tissue that is not of clinical significance.

I demonstrated that the technique can be used to delineate suspicious lesions reliably and that it increases the amount of separation between malignant and benign lesions based on a margin measurement I created (accessed here). It also improved separation between malignant and benign lesions based on average values of the established signal enhancement ratio method. Example segmentations produced by the technique are provided below where each blue square is a 1.75 cm by 1.75 cm patch of tissue containing a radiologically suspicious lesion. Red lines mark the boundary between the suspicious lesion and its surrounding tissue. The samples above the green line represent malignant lesions and below the green line are radiologically suspicious benign lesions.




Note that many of the lesions are extremely small (including malignancies as small as 2 to 3 mm across), a testament to the exceptional MRI enabled breast cancer screening program run at Sunnybrook Hospital in Toronto, Canada where my collaborators are based.


Citation for this paper: