Lightweight CNN Towards Skin Lesions Automated Diagnosis In Dermoscopic Images

Skin cancer is a grave illness that can have devastating consequences. One of the primary diagnostic tools used in its identification is medical imaging, which can provide valuable information about the size, stage, and location of the tumor lesions. This work aims to automate skin lesions diagnosis in dermoscopic images. To succeed this task, a lightweight CNN was developed and compared to various commonly used models based on Tansfer Learning strategy in terms of performance, complexity and inference time. The proposed model was evaluated using the dataset HAM10000 including 10015 dermoscopic images with 8 skin lesions. The reached results indicates that LWCNN is approximately, the most performant, the least complex, and the minor inference time’ model. In fact, the proposed LWCNN reached an accuracy of 0.9889, F1-score of 0.9888, AUC of 0.9968, and a loss of 6.73.10
^-2
. Moreover, its parameters number were 1.612 million and its inference time was 6ms/step. Furthermore, for all the classes, the attained accuracy and F1-score exceeds 0.9894 and 0.9580, respectively. Based on the high performance observed in the study, it is possible that the methods used could serve as a valuable resource for clinicians in examining and diagnosing skin lesions, potentially leading to improved healthcare quality in this domain.