Sunteți pe pagina 1din 1

Towards Early-Stage Malignant Melanoma Detection Using Consumer Mobile Devices

John Breneman
Stanford Center for Professional Development, Department of Electrical Engineering

Abstract:

Digital dermatoscopy techniques for diagnosing malignant melanoma are adapted to consumer-grade mobile devices with macro photography lenses and a prototype application for Android devices is presented.

Malignant Melanoma
Melanoma is the most deadly variety of skin cancer, accounting for the majority of skin cancer related deaths globally Self-examination is proven to be effective for decreasing mortality rates, but there is a large inconsistency in the interpretation of risk factors The ABCDE method for skin self-examination is the most established Endorsed by the American Academy of Dermatologists Three components of the ABCDE method were selected for automated image processing

Preprocessing
Otsu Binarization Small Region Removal
Center

Feature Detection and Description


90

Symmetry axis

Increment counter

45 Symmetry axis

Segmentation
Largest Component Max Size Filter
Do not increment

Lesion
Center

Lesion

Descriptors
Symmetry
Border Region color and texture are variable Use percentage of pixels whose mirrors are included within a region instead Computing a full symmetry map is computationally expensive Calculate symmetry about 0, 45, and 90 degree axes only

Color varies with devices color balancing algorithm Use average gradient magnitude along region border Must use Gaussian blur due to inaccuracies in segmentation Poor lighting prevents absolute scale metrics Use gray-level intensity variance

Color

Flow Diagram
Images courtesy of American Academy of Dermatology

Asymmetry

Border

Color

Dermatoscopy vs. Macro Photography


handyscope by Fotofinder Digital Dermatoscope
Physician centered: Designed for expert users

Results
Printed Phantom Asymmetry Phantom

Future Work
Testing against large datasets
Image segmentation refinement Region matching for change detection over time Lesion location mapping for automatic generation of mole maps

Polarized filters Sub-dermal imaging Existing body of technical literature for automated image processing

Relative ranking of cancerous and non-cancerous lesions assessed by a group of trained dermatologists Asymmetry feature was especially powerful in separating cancer from non-cancer. Sub-sampled symmetry map approach was sufficiently invariant to rotation Border feature varies with focus perhaps Fourier normalization required? Color feature varies with lighting Perhaps rank in 2D hue/chrominance space, not luminance?

Color Phantom

Printing test images does not work pixilation due to printing process Use melanoma phantoms to test individual descriptors Required locally adaptive thresholding due to device shadow in image

Photojojo Macro Lens Adapter for Mobile Devices


Patient

Acknowledgements
Susan Swetter, M.D., David Peng, M.D. Stanford Department of Dermatology Bernd Girod, PhD, Ngai-Man Cheung, PhD. ,Zhi Li - Stanford Electrical Engineering

centered: requires no prior knowledge


Inexpensive,

consumer grade Magnification only Barrel and pincushion distortion Variable imaging conditions

Feature Validation

Mobile Implementation

S-ar putea să vă placă și