Delineation of Skin Strata in Reflectance Confocal Microscopy Images with Recurrent Convolutional Networks
1.
Delineation of SkinStrata in
Reflectance Confocal Microscopy
Images With Recurrent
Convolutional Networks
Alican Bozkurt*, Trevor Gale*, Kivanc Kose, Christi Alessi-Fox, Dana
Brooks, Milind Rajadhyaksha, Jennifer Dy
The Skin &Dermal-Epidermal Junction (DEJ)
• Outer shell of the body
• Exposed to outside effects, e.g. sun,
chemicals, etc...
• Skin cancer is the most common
cancer in USA
• Melanoma is among the deadliest.
https://openi.nlm.nih.gov/detailedresult.php?img=PMC2914370_cde0002-0103-f03&req=4
https://www.cancer.org/cancer/melanoma-skin-cancer/detection-diagnosis-staging/survival-rates-for-melanoma-skin-cancer-by-stage.html
• Most skin cancers emerge from
dermal-epidermal junction (basal
layer)
• Early detection is important
• Early stage→5 Year survival 98%
• Late stage → 5 year Survival 18%
Data & Objective
•Given a stack, label each image Epidermis, DEJ, or Dermis
16.
Classification using singleimages
• Classify images independently (not using
sequence information)
• Inception v3 (Szegedy et al.,2015)
• Trained from scratch
• Other methods
• Hames et al. (2016): BoW + linear SVM
• Kaur et al. (2016): MR8 filters + histogram +
MLP
Hames, Samuel C., et al. "Automated segmentation of skin strata in reflectance confocal microscopy depth stacks." PloS one 11.4 (2016): e0153208.
Kaur, Parneet, et al. "Hybrid deep learning for Reflectance Confocal Microscopy skin images." Pattern Recognition (ICPR), 2016 23rd International Conference on. IEEE, 2016.
17.
Results
• Non-sequential methodscan’t capture structure of the problem
• Epidermis → DEJ → Dermis
Accuracy (%)
Error types
Total #
ErrorEpidermis
→ Dermis
DEJ →
Epidermis
Dermis →
Epidermis
Dermis →
DEJ
Inception-v3 84.87 3 25 8 32 68
Hames et al. 84.48 14 59 11 56 140
Kaur et al. 64.33 32 255 16 99 402
Conclusions
• RCM asbeneficial screening
before biopsy
• Convolutional + recurrent units for
highly structured problem
• Deep learning based methods
beat hand-tuned features
• Adding sequential information
improves performance
• RCNs beat vanilla CNNs
• Using only 3 images gives similar
accuracy to best model
27.
Acknowledgements
• This workwas supported in part by the National Cancer Institute
(NCI) under grants R01CA156773 and R01CA199673, the National
Institute of Biomedical Imaging and Bioengineering’s (NIBIB) Image
Guided Interventions Program under grant R01EB012466, the NCI
Core Center grant P30CA008748, and NIBIB grant R01EB020029.