Skin Lesion Detection with YOLOv3

Yolov3 [70 ] was trained with the ISBI 2017 dataset as the lesion detection part of our system. The dataset was separated as training and validation sets. Then the final system detection performance was evaluated using two different datasets (the ISBI 2017 and the PH2). There are studies about the effectiveness of transfer learning in deep nets [71 (link),72 (link)]. So, in the training phase, we used pretrained weights of ImageNet dataset [73 (link)]. Afterwards, Yolov3 was fine-tuned and re-trained with the skin lesion images. The training parameters of Yolov3 are set as follows: batch size = 64, subdivisions = 16, momentum = 0.9, decay = 0.0005, learning rate = 0.001. Yolov3 was trained through 50,000 epochs and the network weights were saved every 10,000 epochs. Test result showed that the weights saved at 10,000th epoch were the most successful at detecting the location of lesion in the image. The whole implementations and computations were performed on a PC with two Intel Xenon processors, 64 GB RAM, NVIDIA GTX 1080Ti GPU and Ubuntu 14.04 operating system. Python and C programming languages, OpenCv image processing framework were used in the development of the system.