Melanoma Prediction Using CNN Methods. Part 3
CNN Architectures
Convolutional Neural Network architectures are ones of deep learning methods for images and video datasets. Also, CNN architectures are used in healthcare widely.
Regarding the paper (Sarvamangala, D.R. et al., 2021), CNN methods are ones the more used architectures in healthcare and skin lesions analysis. Besides, CNN methods are applied for skin lesions machine learning problems and challenges. Also, CNN architectures are good option when input data can be represented as matrix (Singh, P., 2022). Moreover, Convolutional Neural Networks methods are the most popular instrument for image classification (Mikołajczyk, A. et al., 2018). For instance, the paper (Maia, L., B., 2018) mentions InceptionV3, ResNet50, VGG16 and VGG19 architectures. At the same time, the most popular Deep Neural Networks contains many layers and millions of parameters, which requires to use big datasets (Mikołajczyk, A. et al., 2018). The paper “Evaluation of Melanoma Diagnosis using Deep Features” (Maia, L. et al., 2018) investigate VGG architectures, Inception module (GoogLeNet), and ResNet architectures of Convolutional Neural Networks for mole image classifications for melanoma prediction. Also, the paper (Maia, L. et al., 2018) mention about standard architectures adaption as training from scratch without prepared weights but using random initialization, using pretrained weights, and using other values of other networks for training. The Convolutional Neural Networks part of my research purpose is finding the architecture with best accuracy and prepare the model for using in my product for melanoma prediction. I decided to use the following transfer learning models of Convolutional Neural Networks in my machine learning analysis for my melanoma prediction project as Xception, ResNet50, DenseNet201, VGG16, VGG19, U-Net. These architectures are pretrained using a large-scale hierarchical image dataset (Deng, J. et al., 2009).
With reference to Table 1, HAM10K dataset is one of the most referenced datasets in science papers related to machine learning in melanoma prediction and skin lesions problems, that is why, I reviewed the papers for melanoma prediction using Convolutional Neural Networks based on HAM10000 dataset. The review helps to find better network configuration for my project, also, it allows to verify my results by comparing them with results in the science papers. Table 2 contains my literature review of using Convolutional Neural Networks with HAM10000 dataset. Moreover, most of the CNN architectures, which are reviewed in my work, are pretrained by Convolutional Neural Networks methods. Regarding (Kassani, S. et al., 2019), “pre-training highly reduced the training time of CNN and achieves 85.8% accuracy over 5-classes”.
I sorted records in Table 1 in ascending order by accuracy. The table contains method for classification problems. CNN methods as U-Net are used for segmentation (Ronneberger, O. et al, 2015).
With reference to data in Table 2, I can highlight more often CNN methods as Xception, DenseNet201 and InceptionV3.
HAM10K dataset is one of a set of datasets with photos of moles, which are labeled whether they have melanoma or not. There are many science papers, which use CNN architectures with different datasets. That is why, I collected some science papers of using CNN methods in datasets as ISIC in Table 2. I sorted the records in the Table 2 in ascending order by column accuracy.
Table 2 illustrates that InceptionV3, Xception, VGG16 often used by research in ISIC datasets.
Besides, regarding Table 1 and Table 2, pretrained CNN methods are popular option for deep learning in skin lesions classification problems. As mentioned in (Ha, Q. et al., 2020), one of the wide used practices is using transfer learning approach for CNN models, and after changing last layers, to fine tuning for the specific dataset.
Transfer learning approach is important for medical domain since train data and validation data are limited. As mentioned in (Qureshi, M.N. et al., 2022) “Valuable features are identified by using this pre-trained model even when the training samples are limited”.
The research in (Al-Bander, B. et al., 2021) illustrates good results of InceptionResNetV2 model with accuracy 96.23%.
The benchmark results in (Al-Bander, B. et al., 2021) for 12 CNN architectures as VGG16, VGG19, Inception3, Inception4, InceptionResNet2, ResNet50, DenseNet121, DenseNet169, DenseNet201, Xception, MobileNet and NASNetMobile.
Also regarding (Ha, Q. et al., 2020) one of the winners in SIIM-ISIC 2020 challenge used an ensemble of 18 pretrained EfficientNet models
References
Sarvamangala, D.R. and Kulkarni, R.V., 2021. Convolutional neural networks in medical image understanding: a survey. Evolutionary intelligence, pp.1–22.
[Accessed 05 July 2022]
Singh, P., Kumar, M. and Bhatia, A., 2022, May. A Comparative Analysis of Deep Learning Algorithms for Skin Cancer Detection. In 2022 6th International Conference on Intelligent Computing and Control Systems (ICICCS) (pp. 1160–1166). IEEE.
[Accessed 28 May 2022]
Mikołajczyk, A. and Grochowski, M., 2018, May. Data augmentation for improving deep learning in image classification problem. In 2018 international interdisciplinary PhD workshop (IIPhDW) (pp. 117–122). IEE
[Accessed 25 May 2022]
Maia, L.B., Lima, A., Pereira, R.M.P., Junior, G.B., de Almeida, J.D.S. and de Paiva, A.C., 2018, June. Evaluation of melanoma diagnosis using deep features. In 2018 25th international conference on systems, signals and image processing (IWSSIP) (pp. 1–4). IEEE.
[Accessed 25 May 2022]
Deng, J., Dong, W., Socher, R., Li, L.J., Li, K. and Fei-Fei, L., 2009, June. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition (pp. 248–255). Ieee.
[Accessed 30 May 2022]
Kassani, S.H. and Kassani, P.H., 2019. A comparative study of deep learning architectures on melanoma detection. Tissue and Cell, 58, pp.76–83.
[Accessed 30 May 2022]
Ronneberger, O., Fischer, P. and Brox, T., 2015, October. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention (pp. 234–241). Springer, Cham.
[Accessed 06 July 2022]
Ha, Q., Liu, B. and Liu, F., 2020. Identifying melanoma images using efficientnet ensemble: Winning solution to the siim-isic melanoma classification challenge. arXiv preprint arXiv:2010.05351.
[Accessed 06 July 2022]
Qureshi, M.N., Umar, M.S. and Shahab, S., 2022. A Transfer-Learning-Based Novel Convolution Neural Network for Melanoma Classification. Computers, 11(5), p.64.
[Accessed 28 May 2022]
Al-Bander, B., Yas, Q.M., Mahdi, H. and Al-Hamd, R.K.S., 2021. Benchmarking of deep learning algorithms for skin cancer detection based on ahybrid framework of entropy and VIKOR techniques. Turkish Journal of Electrical Engineering and Computer Sciences, 29(8), pp.2634–2648.
[Accessed 05 July 2022]
