Very deep convolutional networks for skin lesion classification

الملخص EN

medical imaging and diagnostic radiology, is fast becoming a key tool in Computer-Aided Systems.

A visual examination of a skin lesion is one of many potential applications of Computer-Aided Diagnosis (CAD).

In this paper, a classification algorithm was developed and trained based on Densely Connected Convolutional Networks (DenseNets) for skin lesion classification.

This work combines the use of dermatoscopic images pre-processing and deep convolutional network.

Experimental studies were conducted using more than 30,000 dermatoscopic images from multiple open- access archives.

Initially, the paper examined the impact of proposed image pre-processing and tuning on the accuracy of a binary classification CNN.

The result showed that the binary classifier achieves a higher area under the curve (AUC), 0.93, when trained on processed data compared to 0.85 before.

In the second stage, two different data-sets were created a 3-ary classes and 9-ary classes.

The validation results showed that the proposed work achieved validation accuracy of 81.2 ± 1.1% in the 3-ary and of 60.1 ± 1.3%.

In the 9-ary classification studies.

The proposed combination of dermatoscopic images pre-processing and deeper convolutional network can achieve better performance by learning more complex features of the input data with a more efficient memory implementation