Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
Evaluating a deep convolutional neural network for classification of skin cancer
KTH, School of Electrical Engineering and Computer Science (EECS).
KTH, School of Electrical Engineering and Computer Science (EECS).
2018 (English)Independent thesis Basic level (degree of Bachelor), 10 credits / 15 HE creditsStudent thesisAlternative title
Utvärdering av ett deep convolutional neural network för klassificering av hudcancer (Swedish)
Abstract [en]

Computer-aided diagnosis (CAD) has become an important part of themedical field. Skin cancer is a common and deadly disease that a CADsystem could potentially detect. It is clearly visible on the skin andtherefore only images of skin lesions could be used in order to pro-vide a diagnosis. In 2017, a research group at Stanford University de-veloped a deep convolutional neural network (CNN) that performedbetter than dermatologists during classification of skin lesions.This thesis makes an attempt at implementing the method pro-vided in the Stanford report and evaluate the performance of the CNNduring classification of skin lesion comparisons not tested in their study.The previously unseen binary classification use cases are melanomaversus solar lentigo and melanoma versus seborrheic keratosis. Usingtransfer learning, Inception v3 was trained for various skin lesions.The CNN was trained with 16 training classes. During validation ofthe CNN, an accuracy of 68.3% was achieved during a 3-way classi-fication. Testing the same comparisons as the Stanford study an ac-curacy of 71% was achieved for melanoma versus nevus and 91% forseborrheic keratosis versus basal and squamous cell carcinoma. Theaccuracy results for the new comparisons were 84% for seborrheic ker-atosis versus melanoma and 83% for solar lentigo versus melanoma.The results suggest that out of the binary classifications performedin this study, nevus versus melanoma is the most difficult for the CNN.It should be noted that our results were different from the Stanfordstudy and that more statistical methods should have been used whenobtaining the results

Abstract [sv]

Computer-aided diagnosis (CAD) har blivit en viktigt del av det medi-cinska området. Hudcancer är en vanlig och dödlig sjukdom som ett CAD system potentiellt kan upptäcka. Den är klart synlig på hudenoch därför skulle endast bilder av hudskador kunna användas för attge en diagnos. År 2017 utvecklade en forskningsgrupp från StanfordUniversity ett deep convolutional neural network (CNN) som preste-rade bättre än dermatologer vid klassificering av hudskador.

Detta kandidatexamensarbete gör ett försök till att implementerametoden tillhandahållen i Stanford rapporten och utvärdera CNN:etsresultat vid klassifikation av hudskador som inte testades i deras stu-die. De binära fall som tidigare inte har testas är melanoma emot solarlentigo och melanoma emot seborrheic keratosis. Med hjälp av transferlearning tränades Inception v3 för olika hudskador. CNN:et tränadesmed 16 typer av hudförändringar. I valideringsprocessen uppmättesen korrekthet på 68.3% under 3-vals klassifikation. I tester av sammatyp av jämförelser som i Stanford studien uppmätes en korrekthet på71% för melanoma emot nevus och 91% för seborrheic keratosis emotbasal and squamous cell carcinoma. Resultatet av de nya jämförelser-na var 84% för seborrheic keratosis emot melanoma och 83% för solarlentigo emot melanoma.

Resultaten tyder på att av de binära klassificeringarna utförda idenna studie, är nevus emot melanoma den svårast för CNN:et. Detbör noteras att våra resultat skilde sig från Stanford studien och attmer stat

Place, publisher, year, edition, pages
2018. , p. 38
Series
TRITA-EECS-EX ; 2018:220
National Category
Computer Sciences
Identifiers
URN: urn:nbn:se:kth:diva-229740OAI: oai:DiVA.org:kth-229740DiVA, id: diva2:1214297
Subject / course
Computer Science
Educational program
Master of Science in Engineering - Computer Science and Technology
Supervisors
Examiners
Available from: 2018-08-03 Created: 2018-06-06 Last updated: 2018-08-03Bibliographically approved

Open Access in DiVA

fulltext(2357 kB)159 downloads
File information
File name FULLTEXT01.pdfFile size 2357 kBChecksum SHA-512
109e349d08fdab42f951ae8c6e7ae5861f71d426f46df8fa040652673e8af1e89006d24ee0d9b8dae2780f7975576d8de0124b4780877ce5d05408dc3bb1cfd7
Type fulltextMimetype application/pdf

By organisation
School of Electrical Engineering and Computer Science (EECS)
Computer Sciences

Search outside of DiVA

GoogleGoogle Scholar
Total: 159 downloads
The number of downloads is the sum of all downloads of full texts. It may include eg previous versions that are now no longer available

urn-nbn

Altmetric score

urn-nbn
Total: 289 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf