Convolution-Attention Approach for Detecting and Diagnosing of Skin Cancer at An Early Stage
Over the years, diagnostic research work on cancer has continued to increase. In skin cancer, a clear way to know is from apparent skin lesions. The area of the skin becomes darker and scaly. The issues lie in the boundary and edge of these lesions. Some parts of the skin affected might not readily show the signs of skin cancer. To accurately detect the extent of the skin cancer from the lesion boundary remains a difficult task. Precisely knowing the border cut off could mean the difference between accurate and inaccurate treatments. Traditional methods of skin cancer classification require a large amount of such strictly labeled data for training classifiers but the SVM classifier doesn't perform well when we have a large dataset since the necessary training time is longer. Another major limitation of the study is that, the conventional CNN used in the existing study treats all skin cancer features from the image equally, resulting in slow learning and less accuracy. Also, as stated in the existing study, the model was not tested with different epochs, batch sizes, classifiers, and optimizers which resulted in the low accuracy recorded in the study. To address these issues, in this research, we proposed an attention-based convolutional neural network for skin cancer detection. The proposed attention model can focus on vital features of the skin cancer datasets while filtering out a large amount of background noise signals. Based on the literature, this is a more efficient approach with better accuracy. Thus, the main objective of the study is to improve the classification accuracy using an improved deep convolution attention model tested with different epochs, batch sizes, and optimizers. Experimental result in MATLAB 2021a shows that the proposed model attains the best accuracy of (92.51%), precision of (90.76%), F-1 of (93.96%), sensitivity of (96.05%) and specificity of (54.76%). Hence, the proposed system achieved the superior classification accuracy of 92.51%compare to the other classical CNN approaches (DenseNet-201 + SVM, MobileNet + SVM and ResNet-50 + SVM) which achieved accuracy of 86.23%, 85.63% and 85.63 respectively.
Authors
- Ahmed Zainab Tijjani
Publication Details
- Journal
- INTERNATIONAL JOURNAL OF COMPUTER SCIENCE AND MATHEMATICAL THEORY E-ISSN
- Published
- 2026-11-13
- DOI
- https://doi.org/10.56201/ijcsmt.v10.no4.2024.pg70.98
- Primary Topic
- Cutaneous Melanoma Detection and Management
- Type
- article
- Field-Weighted Citation Impact
- 0.00