Jason Hagerty, a doctoral candidate in computer engineering, will defend their dissertation titled “Implication And Applications Of Machine Learning On Biomedical Images.” Their advisor, Dr. Joe Stanley, is a professor in the electrical and computer engineering department.  The dissertation abstract is provided below.

Early detection of cancer yields the best chance for successful treatment. The number of estimated new cases of skin cancer has grown from 87,110 in 2016 [1] to 100,640 in 2024 [2] making it the fifth most diagnosed cancer [3]. Cervical cancer also grown from 12,820 in 2017 [1] to 13,820 in 2024 [3]. The presented research seeks to aid in the early detection of skin cancer, specifically melanoma, and cervical cancer.

Medical imaging ranges in modality including computer tomography, x-ray imaging, digital microscopy, and macro-focus dermoscopy images. The latter two modalities are the focus of the presented work.

Images captured using these modalities are usually highly magnified color images in the RGB color space. To perform a diagnostic evaluation on the captured dermoscopy image, it begins with what is usually a labor-intensive operation that requires a human-in-the-loop to perform the initial segmentation to localizing a region of interest (ROI). Once that segmentation is obtained a physician with years of training and experience will observe biological markers that can used to visually differentiate whether a lesion is benign or malignant. A similar process is used for digital microscopy images. Once a sample has been prepared by slicing and staining a specimen the ROI of the captured image is also localized again a physician then judges the subtle difference in the ROI to deduce a diagnostic score.

All this relies on having image data to use for machine learning as well as creating image processing algorithms. Research in a way to infer how much data is needed given the current state of the art methodologies. Data collection is hampered by the fact that, traditionally, medical data is treated differently than other forms of images and metadata is that a patient is entitled to privacy thus making assembling large medial dataset difficult as well as its release via open source.

The presented work demonstrates methods that machine learning and image processing researchers could leverage to overcome the previously mentioned difficulties with the goal to aid the physician in an effort to increase the number of the cases that can be screened and or evaluated.