Research Report on Melanoma Skin Cancer Detection Using CNN Algorithm and Image Processing

Abstract

Skin cancer is a serious problem that is frequently ignored. When a clinician performs a manual examination, Using imaging data, the human eye might not always be able to correctly identify ailments. In today\'s environment, Multi layered networks are being employed more and more to solve difficulties in our everyday lives. Therefore, we develop an automated computerised system for detecting skin problems using deep neural network approaches. We have incorporated several neural network algorithms into the suggested model and evaluated their results to ascertain the most precise method for recognising the five main skin conditions.CNN, and by using the Keras Sequential API, we have created a new model that achieves an accuracy of roughly 85%. We have since used approaches that make use of pre-trained data to increase accuracy and facilitate comparability. These include the transfer learning models VGG16, RESNET50, and Densenet121. From the set of algorithms utilized in the suggested models, the resnet architecture obtains the highest accuracy of 90%.

Introduction

Skin disorders, while common, are often overlooked—especially in rural and low-resource areas like Bangladesh, India, and Sri Lanka—where access to dermatologists is limited. Many people fail to recognize potentially dangerous conditions like melanoma early, leading to late diagnoses and high mortality rates. Even in advanced countries, timely and accurate detection remains a challenge.

To address this issue, the study proposes an AI-based skin cancer detection system using Convolutional Neural Networks (CNNs), leveraging data from the ISIC archive. The goal is to build an automated image-processing model that accurately identifies skin lesions, particularly melanoma, from dermoscopic images.

Related Work:

Previous research has explored the use of texture analysis, color features, and artificial neural networks (ANNs) for melanoma detection. However, traditional methods often struggle with complex lesion features and segmentation challenges. More recently, deep learning approaches like CNNs have shown superior performance, even rivaling dermatologists in accuracy.

Studies using feed-forward neural networks, Multilayer Perceptrons (MLPs), and ANNs with backpropagation have laid the groundwork for advanced models. CNNs like VGG, ResNet, and DenseNet offer better feature extraction and classification abilities.

Methodology:

• Dataset: The study used the HAM10000 dataset from the ISIC archive, containing thousands of labeled dermoscopic images of seven skin lesion types.

• Preprocessing Steps:

  • Resizing images

  • Normalizing pixel values

  • Label encoding (0–6 for seven lesion types)

  • Data splitting (80% train, 10% test, 10% validation)

  • Data augmentation (rotation, zoom, flip)

• Model Implementation:

  • VGG16: Uses small 3x3 filters, max pooling, and fully connected layers. The model uses the Adam optimizer and categorical cross-entropy loss.

  • ResNet50: Incorporates residual connections to combat vanishing gradients and allows deep training.

  • DenseNet121: Uses dense connections where each layer gets input from all previous layers, improving feature reuse and information flow.

Results & Challenges:

• VGG models (VGG16 and VGG19) showed signs of underfitting—they failed to generalize well, and performance plateaued early despite high training accuracy.

• ResNet and DenseNet offered better architecture designs to combat vanishing gradients and enhance learning.

• CNN-based systems outperform traditional methods by learning complex, hierarchical features directly from raw image data.

Conclusion

This paper discusses the concept of using a CNN model with the Keras Sequential API to detect different types of skin cancer. We have established the kernel size for our convolution and the filter size for our max-pooling. We created the ANN using Keras after flattening these into a 1D single vector. The 10015 images in our database were split into 80:20 test and train sets. 10% validation was applied to the train set. We trained our six-layered CNN model for 50 epochs, obtaining about 79% accuracy with the validation set and 76% accuracy with the test set. We used the CNN method to increase accuracy in VGG11, ResNet50, and DenseNet121 models that utilize ImageNet\'s pre-trained data. Our dataset grew as a result of these models, increasing the model\'s effectiveness. We were able to attain 90% training accuracy with minimal loss using them. For evaluation, we employed 10 epochs for each of these models. We produced a variety of graphics, including faulty prediction graphs, loss and accuracy graphs, and confusion matrices. We also calculated accuracy, supportMetrics such as F1 score and recall were considered. A system that can precisely detect melanoma is critical for dependable outcomes 90% of things just from the information obtained from the picture can be applied in scenarios where this model is more efficient than human detection.In the future, we plan to use our strategies for diagnosing skin cancer in real-world circumstances to support people who faces with restricted usage to medical care.

References

[1] S.S. Han, I.J. Moon, W. Lim, et al. Region-based convolutional 3 neural networks for the identification of keratinocytic skin cancer on the face JAMA Dermatol, 156 (1) (2020), pp. 29-37, 10.1001/jamadermatol.2019.3807 [2] Nazi Z Al, Abir TA. Automatic skin lesion segmentation and melanoma detection: transfer learning approach with U-Net and DCNN-SVM. 2020:371-381. doi:10.1007/978-981-13-7564-4_32. [3] K. Sau, P. Saha Skin cancer classification using an ANN developed with the scaled conjugate gradient algorithm Communications in Computer and Information Science, Vol 1030, Springer Verlag (2019), p. 134, 10.1007/978-981-13-8578-0_11 [4] D. Roffman, G. Hart, M. Girardi, C.J. Ko, J. DengA Multi-Parameterized ANN approach for Predicting benign Skin Cancer Sci Rep, 8 (1) (2018), 10.1038/s41598-018-19907-9 [5] Nagavkar G, Potdar K. A Sign language translator glove view project blind help system view project Kedar Potdar NVIDIA language recognition system. A classifier employing neural network architecture based. [6] The American system for recognising sign language. Vol 5.; 2017. https://www.researchgate.net/publication/319624727 [7] M. A. Sheha, M. S. Mabrouk, A. Sharawy,et al., In the study ‘Automatic Detection of Melanoma Skin Cancer Using Texture Analysis,’ published in the International Journal of Computer Applications, Vol. 42, no. 20, pp. 22–26, 2012. [8] V. Noronha, U. Tsomo, A. Jamshed, M. Hai, S. Wattegama, R. Baral, M.Piya, and K. Prabhash, “A fresh look at oncology facts on South central Asia and saarccountries”,South Asian journal of cancer, vol. 1, no. 1, p. 1, 2012.

Copyright

Copyright © 2025 P. S. Suma, Mrs. Sharvani V. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.