template matching for detection of starry milia-like cysts in dermoscopic images
DESCRIPTION
A poster explaining the work done in detecting Starry Milia like Cysts in Dermoscopic images using template matching techniquesTRANSCRIPT
Department of Electrical and Computer Engineering
Introduction
Results and Conclusions
Methods
Milia-like cysts are small white or ivory round/oval structures seen on dermoscopy images, from a device that emphasizes deep structures). MLCs are categorized into starry (< 1/3 mm diameter) and cloudy ( > 1/3 mm diameter). MLCs are often found in seborrheic keratosis (SK), the most common benign lesion. An example of both types of MLCs within an SK can be seen here.
Starry MLCs (stars) have a sensitivity of 90.5% and specificity of 45.7% for seborrheic keratosis.1
MLCs, if detectable, can aid in distinguishing benign lesions frommalignant lesions, such as melanoma.
MLC mimics—scales, bubbles, and skin pores, must be removed to achieve automatic MLC detection.
Multiple template correlations are needed to optimally characterize fine structures 3 correlations used unprocessed ROIs, shape-limited ROIs, and histogram-equalized ROIs. A template-matching method gives a useful, quantized, size determination For ROIs with ill-defined borders, shape can be determined based on central ROI statistics Central statistics are also key in MLC versus mimic decision Starry MLCs can be distinguished from close mimics with acceptable accuracy Fine ROI features and ROI neighborhoods are important in medical imaging. .
1. Pre-processingLesion borders were manually drawn, bubble and hairs were removed
2. Lesion divided into 13x13 blocksLesion area was segmented into 13x13 blocks. The brightest pixel in each block was chosen as the center of a candidate starry MLC (star). Stars can vary in size, therefore only the 11x11 center of stars were analyzed to allow for a measure of consistency (Figure 2).
3. Candidate stars selectionA 3x3 mean filter was applied to reduce image noise. The mean of the lesion intensity was calculated. Candidate stars were removed if they did not exceed the threshold: 0.85*(lesion mean intensity).
4. Correlation coefficient—Template matchingCorrelation γ(x,y) with star templates created by characterizing pre-selected example starry MLCs:
where x = 0, 1, 2, 3 . . ., M-1, y = 0, 1, 2 . . ., N-1, MxN is the size of f, is the mean value of pixels in w and is the mean value of the region that is coincident with the current location of w.2 The correlation coefficient was scaled to between -1 to 1, with 1 having the strongest correlation. Starry MLCs can vary in size, therefore three differently sized templates were used: 11x11, 17x17, and 23x23.5. Application of Range as a threshold
All starry MLCs have a bright center pixel and radially decreasing intensity. Therefore, we used the Range of the maximum and minimum pixel intensity values within the 11x11 center. Those that met the threshold: 0.14, 0.075, and 0.03 (scaled to between 0 and 1), respectively, continued for further processing.
Figure 5: ROC for MLC detection--SK vs. melanoma.
References1. Stricklin, S.M., Stoecker, W.V., ..& Mahajan, S.K. (2011). Cloudy and starry milia-like cysts: How well do they distinguish
seborrheic keratoses from malignant melanoma J European Academy of Dermatology and Venereology, 25(10), 1222-4.2. Gonzalez, R.C. & Woods, R.E. (2002). Digital Image Processing. 2nd ed. New Jersey: Prentice Hall, pp. 698-704.
Analysis and Design
6. Star shape analysisShape features (e.g. area, perimeter, roundness, solidity, major and minor axis, eccentricity, and major axis/equivalent diameter) and the second correlation coefficient between the candidate star and the ideal star template are then calculated.
SAS Logistic Regression Results
Lesion borders drawn, hairs and bubbles removed
Lesion divided into 13x13 blocks
Candidate star selection: 0.85* (Lesion mean intensity)
Viswanaath Subramanian1, Randy H. Moss1, Ryan K. Rader2, Sneha K. Mahajan1, William V. Stoecker1,2
1Dept. of Electrical and Computer Engineering, Missouri University of Science & Technology, Rolla, MO USA,
2Stoecker & Associates, Rolla, MO USA
Template Matching for Detection of Starry Milia-like Cysts in Dermoscopic Images
1st Correlation coefficient: star templates
Threshold from star center intensity range
Candidate star shape analysis
Logistic regression analysis
Figure 1: SK with MLCs
Figure 2: Starry MLC ROI detail, red circle: central pixel, black rim: 11x11 region for range determination
Parameter Chi-square Pr > Chi-sq
Star size 24.69 <.0001
First correlation coefficient with ideal star template
29.17 <.0001
Intensity difference between star and surround
25.60 <.0001
Mean of 11x11 star center 8.68 0.0032
Rise time of 11x11 star center
30.83 <.0001
Correlation coefficient to the star shape template
11.19 0.0008
Equalized correlation coefficient
29.24 <.0001
7. Notes on methods. Region of interest (ROI): star MLC candidate Block size 13x13 optimized: ↑ size ↑ ⇒
computational efficiency, ↓ size ↓ loss from ⇒dupl. MLCs
Lesion size determined by correlation with 3 templates: 11x11, 17x17, 23x23
3 thresholds needed to ↓ false positives: lesion mean threshold, duplicate removal, and range threshold
Star shape determined after 3x3 average filter smoothing: area pixels with intensity within 10% of smoothed object intensity
2nd Correlation coefficient: shape templates
Figure 3: Formation of shape of the star for a true starry MLC. (a) Original image – Seborrheic Keratosis. (b) Lesion and bubble masks applied on blue plane. (c) Enlarged version of a starry MLC. (d) Histogram equalized starry MLC. (e) Enlarged version of smoothed starry MLC. (f) Initial shape of the star. (g) Thinned – one pixel width shape of the star. (h) Final filled shape of the star shifted to the center.