Minimizing false negative rate in melanoma detection and providing insight into the causes of classification
Authors:
Ellák Somfai,
Benjámin Baffy,
Kristian Fenech,
Changlu Guo,
Rita Hosszú,
Dorina Korózs,
Fabrizio Nunnari,
Marcell Pólik,
Daniel Sonntag,
Attila Ulbert,
András Lőrincz
Abstract:
Our goal is to bridge human and machine intelligence in melanoma detection. We develop a classification system exploiting a combination of visual pre-processing, deep learning, and ensembling for providing explanations to experts and to minimize false negative rate while maintaining high accuracy in melanoma detection. Source images are first automatically segmented using a U-net CNN. The result o…
▽ More
Our goal is to bridge human and machine intelligence in melanoma detection. We develop a classification system exploiting a combination of visual pre-processing, deep learning, and ensembling for providing explanations to experts and to minimize false negative rate while maintaining high accuracy in melanoma detection. Source images are first automatically segmented using a U-net CNN. The result of the segmentation is then used to extract image sub-areas and specific parameters relevant in human evaluation, namely center, border, and asymmetry measures. These data are then processed by tailored neural networks which include structure searching algorithms. Partial results are then ensembled by a committee machine. Our evaluation on the largest skin lesion dataset which is publicly available today, ISIC-2019, shows improvement in all evaluated metrics over a baseline using the original images only. We also showed that indicative scores computed by the feature classifiers can provide useful insight into the various features on which the decision can be based.
△ Less
Submitted 9 March, 2021; v1 submitted 18 February, 2021;
originally announced February 2021.
