The effect of quantization at different resolutions on the classification of lesion texture in dermoscopic images using support vector machine. Issue 6 (27th May 2022)

• Record Type: Journal Article
• Authors: Mahmoud‐Ghoneim, Doaa
• Is Part Of: International journal of imaging systems and technology. Volume 32:Issue 6(2022)
• Page Start: 1953
• Page End: 1962
• Publication Date: 2022-05-27
• Subjects: automated diagnosis -- cooccurrence matrix -- melanoma -- skin cancer -- support vector machine -- texture analysis; Imaging systems -- Periodicals; Image processing -- Periodicals; 621.367
• DOI: 10.1002/ima.22759
• Languages: English
• ISSNs: 0899-9457

Abstract

Abstract: Dermoscopic images have been recently used for automated diagnosis of skin lesions with successful results and a potential for widescale applications. Cooccurrence matrix (COM) is a statistical method of texture feature extraction that remains possibly the most interpretable and expandable method in its category. COM is investigated on four groups of dermoscopic images (high resolution ROI: HRR, medium resolution ROI: MRR, low resolution ROI: LRR, and mixed resolution: MixR), using three different gray level quantization values ( b = 25, 15, and 8). Three skin regions were outlined: melanoma, common nevi, and normal skin. Texture features were extracted then Support Vector Machine (SVM) was used as a machine learning classification algorithm. Sensitivity, specificity, and the area under the ROC curve (AUC), were calculated and compared. In the classification processes that involved either melanoma or common nevi against normal skin, MRR at quantization value b = 15 showed an effective and a high‐performance choice. On the other hand, when classification involved the two lesions, melanoma, and common nevi, quantization value b = 25 showed best results at both HRR and MRR. A high quantization value at intermediate resolution shows to be an effective choice that balances performance with calculation cost. In all cases, quantization at b = 8 deteriorated the classification output even at high resolution. MixR showed good results when normal skin was classified against either type of lesion. In conclusion, selecting the appropriate quantization is a crucial factor that outweighs resolution effect by itself, and can significantly alter the SVM‐COM model performance in automated skin cancer diagnosis.