THE USE OF RESNET50 FOR SKIN CANCER ANALYSIS

Dr Hema N, M Satya Poornima, Madhurya R, Pragya, Pravallika K

Department of Information
Science RNS Institute of
Technology, Bengaluru
[Link] poornima@[Link]
[Link]@[Link]
m [Link]@[Link]

[Link]@[Link]

II. LITERATURE SURVEY

Abstract: Skin cancer remains a critical global health concern,
[1]. The study uses refined deep learning models,
with over 2.1 million cases diagnosed annually, many in areas
specifically DenseNet-121, to analyze the classification of
with limited access to dermatological care. Providing an
skin cancer. Data was sourced from HAM10000 and ISIC
accurate diagnosis on time is essential but challenging in rural
datasets, comprising 3297 RGB images labeled as cancerous
and backward areas. The growth of artificial intelligence (AI)
or non- [Link] machine learning (CNN, SVM,
and deep learning has shown significant potential in aiding skin
Random Forest) and adapted learning models
cancer detection and classification. This study focuses on using
(EfficientNetB0, ResNet34, VGG16, Inception version 3,
deep learning models like CNN for categorization of skin and DenseNet121) were evaluated. DenseNet-121, with
[Link] review discusses different CNN architectures and additional dense layers and a sigmoid activation function,
methodologies, emphasizing the need for further innovation to yielded the highest accuracy at 87%. The study concludes
enhance model accuracy across multiple classes, thus that fine-tuned DenseNet-121 is highly suitable for grouping
supporting widespread, accessible diagnostic solutions. skin diseases into different categories, enhancing diagnostic
precision, but recommends subsequent research with more
keywords: AI (ARTIFICIAL INTELLIGENCE), CNN diverse skin color data for better generalization.
(CONVOLUTIONAL NEURAL NETWORKS), API [2]. This study utilized 2,357 images from the ISIC
(APPLICATION PROGRAM INTERFACE), HAM10000 database, featuring nine skin cancer types, and applied pre-
(HUMAN AGAINST MACHINE WITH 10000 IMAGES) , processing approaches. Deep neural model, specifically
VGG16 (VISUAL GEOMETRY GROUP). ResNet50, were employed for image partition, attribute
extraction, and classification. By leveraging these methods,
I. INTRODUCTION the model attained an accuracy of 91.32%, outperforming
Skin cancer classification has become a critical area of research previous models. This research underscores the efficacy of
on account of the rising incidence of cases globally. CNNs in skin cancer detection, demonstrating the capability
Conventional diagnostic techniques are reliant on dermatological of automated diagnostic tools to enhance early cancer
expertise, which can be limited in availability, especially in detection effectiveness and suggesting that future models
outlying regions. CNN has emerged as an effective technique for could further refine accuracy by incorporating diverse
image-based classifications and shows great promise in datasets and improved feature extraction.
dermatology for analyzing skin lesions. CNNs can automatically [3]. The study developed a pioneering deep CNN structure,
learn and detect complex features within medical images, SkinLesNet, aimed at differentiating skin tumor types and
enabling efficient and accurate categorization of skin cancer detecting melanoma. The researchers used the PAD-UFES-
types without manual feature extraction. 20-Modified dataset, containing 1,314 smartphone images of
three skin lesion types: melanoma, nevus, and seborrheic
The goal of this project is to develop a CNN-based model that keratosis. The methodology included data preprocessing,
can distinguish between malignant and benign forms of skin augmentation, and training on Google Colab using
cancer. By training the model on a diverse set of dermoscopic TensorFlow and Keras. SkinLesNet's performance was
images, this project seeks to leverage CNN’s capability to evaluated and compared to established models like
identify subtle distinctions between malignant and benign ResNet50 and VGG16, achieving a 96% accuracy, notably
lesions, thus aiding in more accurate diagnoses. The outcome of outperforming both. The study concluded that SkinLesNet
this project could potentially support healthcare professionals in could provide an accessible and effective tool for early
making faster and more reliable assessments, enhancing early melanoma detection.
detection efforts, and improving patient care outcomes in [4].The paper presents a unique approach to skin cancer
dermatology.
 evaluation using a combination of feature-extracting neural [8] The application of neural networks to image-based skin
network and discrete wavelet transformation (DWT) to cancer diagnosis is examined in this [Link] the
enhance classification accuracy. The study employs the ISIC 2018 dataset with 11,527 images, the study applied
HAM10000 dataset, focusing on the preprocessing of skin CNN frameworks ,including ResNet50, Inception Version
disease images, DWT-based feature extraction, and CNN- 3, and Inception ResNet, with image preprocessing via
driven classification to boost the early detection of cancerous ESRGAN for quality enhancement. Each model was fine-
and non-cancerous lesions. By identifying critical features in tuned for optimal accuracy, and the InceptionV3 model
images using DWT and applying CNNs for classification,The attained top- most accuracy level of 85.7%. This research
suggested model outperformed traditional models like underscores the impression of hierarchical learning in
artificial neural architecture (ANN) in classification, with a accurate and automated skin cancer detection, though
sensitivity of 94% and specificity of 91%., significantly further improvement on larger datasets is recommended
surpassing conventional models such as artificial neural for real-world applicability.
architecture (ANN) and multilayer perceptrons with respect to [9]. This paper presents an improved network, named
accuracy. This method provides a robust framework for SkinTrans, for segregating skin cancer images. The method
clinical application, demonstrating that CNNs combined with leverages Vision Transformers (ViT) with multi-scale and
DWT can offer fast and accurate evaluations, supporting early overlapping sliding windows, enhancing image feature
diagnosis and better treatment outcomes. While the model’s extraction through multi-scale patch embedding.
efficacy is promising, additional refinements are Contrastive learning is applied to achieve similar encoding
recommended for computational efficiency to better suit real- for like data, maximizing inter-class differences. Testing
time clinical applications. on the HAM10000 dataset and a clinical dataset achieved
[5]. This study examined skin cancer grouping leveraging 94.3% and 94.1% accuracy, respectively, demonstrating
fusion of various models, combining DenseNet-201, SkinTrans's effectiveness. Results suggest this model could
MobileNet, and ResNet-50 with a support vector model. Using aid dermatologists in accurate diagnosis, though an
the World Symposium on Biomedical Imaging 2016 dataset, extended research on clinical experiments are necessary to
models were trained on augmented dermoscopy images of validate its robustness.
benign and melanoma lesions. The methodology incorporated [10].This study offers a methodical evaluation of
transfer learning from pre-trained CNNs and applied an SVM convolutional neural network applications in the
classifier for enhanced classification. The hybrid model categorization of skin cancer., comparing the functionality
combines DenseNet-201 and MobileNet accomplishing the of AI-based models with human experts. It delineates the
greatest precision of 88.02%. The results suggested that hybrid capability of CNNs to classify skin cancers, including
CNN-SVM models improve classification accuracy, melanoma, with accuracy comparable to or better than
suggesting future improvements with diverse datasets and dermatologists in controlled settings. The review assesses
parameter tuning. CNN effectiveness on dermoscopic, clinical, and
[6]. This paper utilized CNN-based techniques to group histopathological images, noting that CNNs generally
skin cancer into various types , leveraging the ISIC perform well in binary and multiclass tasks. However, the
dataset with 3,297 non-cancerous and cancerous study requires additional research in real-world clinical
[Link] an 80-20 split for instruction and settings and on diverse populations to enrich the
assessment, and extensive hyperparameter tuning, the generalizability and clinical relevance of CNN-based tools.
CNN achieved a 98.5% accuracy, significantly [11]. This paper uses a two-stage model using CNN and
outperforming DNN models. The CNN's convolutional SVM was implemented to classify melanoma and
squamous cell carcinoma. Leveraging the U-Net for lesion
layers, which effectively extract essential features from
segmentation and CNNs such as ResNet-152 for
targeted regions, allowed the model to reduce parameters
classification, the model achieved up to 90% accuracy,
and improve classification accuracy and speed. This highlighting the possibilities of transfer learning and deep
study concludes that CNNs offer promising precision learning in medical evaluation. Future work proposed
and efficiency in in clinical imaging for skin cancer includes expanding model training with unscaled images to
identification. improve detection across diverse skin conditions.
[7]. This paper presents a comprehensive review of recent [12].This paper presents a methodology for identifying and
advancements in analysis of diverse categories using deep labeling skin cancers using image analysis andlabelling
learning. The review identifies common deep learning models algorithms. Details from the ISIC 2019 dataset were used,
and methodologies such as convolutional neural networks consisting of 800 dermoscopic images. Key tools include
(CNNs), transfer learning, and ensemble methods. The authors the Dull Razor method, Gaussian and Median filters for
evaluate these techniques across various datasets like preprocessing, and k-means clustering for segmentation.
HAM10000 and ISIC, highlighting issues related to dataset bias Feature extraction employed the ABCD method and
and the necessity for improved model generalization. The paper GLCM, whereas classification was performed using a
concludes that despite the advancements in architectures Multi-class Support Vector Machine (MSVM), achieving
diagnostic accuracy significantly,further work is needed on 96.25% accuracy. This approach demonstrates a robust
model generalizability and hardware compatibility for real-time pipeline for early identification of skin cancer,enhancing
clinical application. diagnostic precision.
 III. OBJECTIVES generalization across diverse datasets and real-world
1. The project aims to develop a mechanism that applicability, particularly in handling variations in skin tones
efficiently detects skin cancer from images, leveraging and achieving reliable performance in clinical settings. This
deep learning techniques. survey highlights the promise of deep learning for reliable
2. Key goals include ease of use, global accessibility, and accessible skin cancer diagnostics, while also
and high accuracy with a low false positive rate. emphasizing the prominence of improving dataset diversity,
3. Machine learning advancements in image model robustness, and clinical validation for broader
recognition and visual data processing to support reliable implementation.
skin cancer detection.
4. The scope covers ongoing research, algorithm VI. REFERENCES
development, and early diagnosis, which can greatly [1].A. Bello, S.-C. Ng, and M.-F. Leung, "Skin Cancer
improve treatment outcomes. Classification Using Fine-Tuned Transfer Learning of
5. AI-based image analysis aids dermatologists in DENSENET-121," Applied Sciences, vol. 14, no. 17, pp.
making precise and early diagnosis of skin cancer. 7707, Aug. 2024

[2]. C. Kavitha, S. Priyanka, M. Praveen Kumar, and V.

IV. PROPOSED SYSTEM Kusuma, "Skin Cancer Detection and Classification using
Our approach focuses on leveraging CNN to effectively Deep Learning Techniques," Procedia Computer Science,
categorize skin cancer pictures that are first altered into an array vol. 235, pp. 2793-2802, 2024.
with a 32x32 pixel resolution. The RGB color channels are
separated and stored individually in a CSV file, a procedure that [3]. A. Muhammad, K. Kiani, T. Mansouri, and N. Topping,
entails breaking down the image into its component elements for "SkinLesNet: Classification of Skin Lesions and Detection of
subsequent analysis. Images are preprocessed through resizing, Melanoma Cancer Using a Novel Multi-Layer Deep Network,"
normalization, and augmentation to enhance model [Link] Cancers, vol. 16, no. 1, pp. 1-15,2024
last,the outputs are classified into benign and malignant types of
skin cancers. [4]. S.P.A. Claret, J.P. Dharmian, and A.M. Manokar, "Artificial
Intelligence-driven Enhanced Skin Cancer Diagnosis:
Leveraging Convolutional Neural Networks with Discrete
SYSTEM ARCHITECTURE Wavelet Transformation," Egyptian Journal of Medical Human
Genetics, vol. 25, no. 50, pp. 1-12, 2024.

[5] . D. Keerthana, V. Venugopal, M. K. Nath, and M.

Mishra, "Hybrid Convolutional Neural Networks with SVM
Classifier for Classification of Skin Cancer," Biomedical
Engineering Advances, vol. 5, no. 100069, pp. 1-8, 2023

[6].Albawi, S., Arif, M. H., and Waleed, J., "Skin cancer

classification dermatologist-level based on deep learning
model," Acta Scientiarum. Technology, vol. 45, no. 1, pp.
e61531, May 2023.

[7]. M. Naqvi, S. Q. Gilani, T. Syed, O. Marques, and H.-C.

Kim, "Skin Cancer Detection Using Deep Learning—A
Review," Diagnostics, vol. 13, no. 1911, pp. 1–26, doi:
10.3390/diagnostics13111911,May 2023.

[Link] [8] .W. Gouda, N. U. Sama, G. Al-Waakid, M. Humayun, and

There have been major improvements in classification of skin N. Z. Jhanjhi, "Detection of Skin Cancer Based on Skin
cancer types using deep learning, especially with Lesion Images Using Deep Learning," Healthcare, vol. 10,
Convolutional Neural Networks, transfer learning models, and no. 1183, pp. 1–18, doi: 10.3390/healthcare1007118,2022.
hybrid approaches. Studies consistently demonstrate that
CNN-based models, such as DenseNet-121, ResNet, and [9]. C. Xin, Z. Liu, K. Zhao, L. Miao, Y. Ma, X. Zhu, Q.
specialized architectures like SkinLesNet, can attain high Zhou, S. Wang, L. Li, F. Yang, S. Xu, and H. Chen, "An
precision in binary and multiclass skin cancer division with improved transformer network for skin cancer
some models surpassing traditional methods. Methods such as classification," Computers in Biology and Medicine, vol.
data amplification, preprocessing, and integration of SVM 149, pp. 105939, doi:
classifiers and discrete wavelet transformations have further 10.1016/[Link].2022.105939,2022.
improved diagnostic performance. However, challenges
remain in model
[10].S. Haggenmüller, R. C. Maron, A. Hekler, J. S. Utikal,
C. Barata, R. L. Barnhill, et al., "Skin cancer
classification via convolutional neural networks: systematic
review of studies involving human experts," Eur. J. Cancer,
vol. 156, pp. 202-216, doi: 10.1016/[Link].2021.06.049,Sep-
2021.

[11].Pushpalatha, A., Dharani, P., Dharini, R., & Gowsalya, J.

,Skin Cancer Classification Detection using CNN and SVM.
Journal of Physics: Conference Series, 1916, 012148,2021.

[12]. M. K. Monika, N. A. Vignesh, C. U. Kumari, M. N. V.

S. S. Kumar, and E. L. Lydia, "Skin Cancer Detection and
Classification Using Machine Learning," Materials Today:
Proceedings, doi: 10.1016/[Link].2020.07.366,Aug 2020.