Our paper “Keratinocytic Skin Cancer Detection on the Face using Region-based Convolutional Neural Network” was published on JAMA Dermatology. To my knowledge, the performance of cancer detection was compared with that of dermatologists for the firs..
Deep Learning/Papers2read 2019. 12. 5. 13:14Our paper “Keratinocytic Skin Cancer Detection on the Face using Region-based Convolutional Neural Network” was published on JAMA Dermatology. To my knowledge, the performance of cancer detection was compared with that of dermatologists for the first time in dermatology. Because most of previous studies were classification studies, preselection of end-user was essential. In addition, there were numerous false positives because training data set did not include enough number of common disorders and normal structures.
With the assistance of R-CNN, we trained neural networks with 1,106,886 image crops to localize and diagnose malignancy. The algorithm detects suspected lesion and shows malignancy score and predicts possible diagnosis (178 disease classes).
We used region-based CNN (faster-RCNN; backbone = VGG-16) as a region proposal module, and utilized CNN (SE-ResNet-50) to choose adequate lesion, and utilized CNN (SE-ResNeXt-50 + SENet) to determine malignancy. We chose a multi-step approach to reduce the dimension of problem (object detection -> classification).
The AUC for the validation dataset (2,844 images from 673 patients comprising 185 malignant, 305 benign, and 183 normal conditions) was 0.910. The algorithm’s F1 score and Youden index (sensitivity + specificity - 100%) were comparable with those of 13 dermatologists, while surpassing those of 20 non-dermatologists (325 images from 80 patients comprising 40 malignant, 20 benign, and 20 normal). We are performing an additional work with large scale external validation data set. The pilot result is similar with this report, so I hope I will submit soon.
Web DEMO (https://rcnn.modelderm.com) of the model is accessible via smartphone or PC, to facilitate scientific communication. Sorry for the slowness of the DEMO because it runs on my personal computer despite of the multi-threading and parallel processing with 2080 x1 and 1070 x1.
Thank you.
Paper : https://jamanetwork.com/journals/jamadermatology/article-abstract/2756346
Screenshot : https://i.imgur.com/2TCkdHf.png
Screenshot : https://i.imgur.com/IEZLfOg.jpg
DEMO : https://rcnn.modelderm.com
https://m.facebook.com/groups/107107546348803?view=permalink&id=1021762028216679&sfnsn=mo
With the assistance of R-CNN, we trained neural networks with 1,106,886 image crops to localize and diagnose malignancy. The algorithm detects suspected lesion and shows malignancy score and predicts possible diagnosis (178 disease classes).
We used region-based CNN (faster-RCNN; backbone = VGG-16) as a region proposal module, and utilized CNN (SE-ResNet-50) to choose adequate lesion, and utilized CNN (SE-ResNeXt-50 + SENet) to determine malignancy. We chose a multi-step approach to reduce the dimension of problem (object detection -> classification).
The AUC for the validation dataset (2,844 images from 673 patients comprising 185 malignant, 305 benign, and 183 normal conditions) was 0.910. The algorithm’s F1 score and Youden index (sensitivity + specificity - 100%) were comparable with those of 13 dermatologists, while surpassing those of 20 non-dermatologists (325 images from 80 patients comprising 40 malignant, 20 benign, and 20 normal). We are performing an additional work with large scale external validation data set. The pilot result is similar with this report, so I hope I will submit soon.
Web DEMO (https://rcnn.modelderm.com) of the model is accessible via smartphone or PC, to facilitate scientific communication. Sorry for the slowness of the DEMO because it runs on my personal computer despite of the multi-threading and parallel processing with 2080 x1 and 1070 x1.
Thank you.
Paper : https://jamanetwork.com/journals/jamadermatology/article-abstract/2756346
Screenshot : https://i.imgur.com/2TCkdHf.png
Screenshot : https://i.imgur.com/IEZLfOg.jpg
DEMO : https://rcnn.modelderm.com
https://m.facebook.com/groups/107107546348803?view=permalink&id=1021762028216679&sfnsn=mo
