Article: A fine tuning approach using modified DenseNet model for skin cancer classification Journal: International Journal of Medical Engineering and Informatics (IJMEI) 2023 Vol.15 No.4 pp.323 - 335 Abstract: Skin cancer is one of the deadliest types of all cancers and melanoma is the severest form of it. In this paper, we have used three state-of-the-art deep convolutional neural network (CNN) models such as DenseNet121, DenseNet169, DenseNet201 and proposed modified-DenseNet121, modified-DenseNet169, modified-DenseNet201 to detect and classify the different forms of skin cancer more accurately. The modifications are done by replacing the top layers with one global average pooling layer, five consecutive fully connected layers, 50% dropout in all fully connected layers and finally, one softmax layer allowing to classify into seven class labels. Further, we have adopted data augmentation and fine-tuning approach to improve the classification performance. A comparative performance analysis of the original DenseNet and the modified-DenseNet models has been presented. To evaluate our model performances, we have used the HAM10000 dataset. We have found that the modified-DenseNet201 model has achieved the most superior result, which is 95.07% overall accuracy and 97.30% melanoma accuracy whereas, the original DenseNet201 model has achieved 81.29% overall accuracy and 78.38% melanoma accuracy. Our findings can offer a second opinion to dermatologists in clinical diagnosis. Inderscience Publishers - linking academia, business and industry through research

Title: A fine tuning approach using modified DenseNet model for skin cancer classification

Authors: Ayan Mondal; Vimal K. Shrivastava

Addresses: School of Electronics Engineering, Kalinga Institute of Industrial Technology (KIIT), Bhubaneswar, India ' School of Electronics Engineering, Kalinga Institute of Industrial Technology (KIIT), Bhubaneswar, India

Abstract: Skin cancer is one of the deadliest types of all cancers and melanoma is the severest form of it. In this paper, we have used three state-of-the-art deep convolutional neural network (CNN) models such as DenseNet121, DenseNet169, DenseNet201 and proposed modified-DenseNet121, modified-DenseNet169, modified-DenseNet201 to detect and classify the different forms of skin cancer more accurately. The modifications are done by replacing the top layers with one global average pooling layer, five consecutive fully connected layers, 50% dropout in all fully connected layers and finally, one softmax layer allowing to classify into seven class labels. Further, we have adopted data augmentation and fine-tuning approach to improve the classification performance. A comparative performance analysis of the original DenseNet and the modified-DenseNet models has been presented. To evaluate our model performances, we have used the HAM10000 dataset. We have found that the modified-DenseNet201 model has achieved the most superior result, which is 95.07% overall accuracy and 97.30% melanoma accuracy whereas, the original DenseNet201 model has achieved 81.29% overall accuracy and 78.38% melanoma accuracy. Our findings can offer a second opinion to dermatologists in clinical diagnosis.

Keywords: CNN; DenseNet121; DenseNet169; DenseNet201; dermoscopy; fine-tuning; melanoma; skin cancer.

DOI: 10.1504/IJMEI.2023.132612

International Journal of Medical Engineering and Informatics, 2023 Vol.15 No.4, pp.323 - 335

Received: 27 Feb 2021
Accepted: 25 May 2021
Published online: 30 Jul 2023 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article

Title: A fine tuning approach using modified DenseNet model for skin cancer classification

Keep up-to-date