Investigating Skin Cancer Diagnosis with Machine Learning - A Comparative Study of Feature Selection and Feature Extraction Techniques

Aug 6, 2026

Upload by Violet

Early skin cancer detection improves treatment outcomes, motivating automated image-based classification. This project evaluates how feature selection and feature extraction affect model performance for skin cancer image classification on the HAM10000 dataset. Images are resized and converted to grayscale for non-deep approaches, then dimensionality reduction is applied using resizing, Lasso-based selection, Select K Best, and PCA. Deep and transfer learning are compared using VGG16, Xception, InceptionV3 plus LightGBM, and a fine-tuned transfer learning pipeline. Binary and multi-class results are assessed with AUC and accuracy using stratiﬁed 3-fold cross-validation. Fine-tuning yields the best overall accuracy, while AUC may drop due to overfitting.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages1

Reading Time3 min

AccessFree to download

Introduction

Objectives

Methods

Results

Summary

Future Work

References

Honors Project Presentation Department of Computer Science May 3, 2023 Alex Huynh Project Advisor: Man Basnet Investigating Skin Cancer Diagnosis with Machine Learning: A Comparative Study of Feature Selection and Feature Extraction Techniques Introduction Skin cancer is a common cancer that requires early detection for eﬀective treatment. Machine learning, particularly deep learning, has shown potential in detecting and classifying skin cancer images. However, high image dimensionality challenges classiﬁcation models to learn relevant information. To address this, feature selection/extraction techniques reduce image dimensionality. Moreover, machine learning models can learn from more relevant data and capture less noise. This study investigated the performance of various feature selection/extraction techniques for skin cancer image classiﬁcation on the HAM10000 dataset. We classiﬁed the images using three popular binary and multi-class classiﬁcation models. Lastly, we compared our results against a ﬁne-tuned transfer learning model. Objectives ● Determine how changing the number of features of a skin cancer image aﬀects the classiﬁcation accuracy and AUC. ● Determine which feature selection/extraction techniques, when applied to skin cancer images, yield high classiﬁcation accuracy and AUC. Results Methods 1. Reduce initial complexity: transform all images to grayscale and reduce all images from 600x450 to 300x225 pixels. 2. Employ dimensionality reduction: a . Feature selection i. Resizing-Reduce image dimensionality to 10%-100% of the original size. ii. Lasso-Perform with varying alpha values. iii. Select K Best-Apply on "74%" sized images (dimensions of 224x224) to compare with transfer learning. b. Feature extraction i. Principal component analysis (PCA) -Compare over variance explained on the 74% sized images. ii. Pre-trained deep learning models VGG16, Xception, and InceptionV3 (on RGB images) -Remove the fully connected layer and feed the output into the LightGBM classiﬁer. iii. Utilize a ﬁne-tuned transfer learning approach (on RGB images), which involves freezing the base, training, unfreezing the base, andre-training. 3. Perform two classiﬁcation tasks: binary classiﬁcation and multi-class classiﬁcation. a . Measure area under the curve (AUC), binary accuracy, and multi-accuracy for each task. b. Resizing was performed on three classiﬁers: LightGBM, XGBoost, and CatBoost. Due to similar performance between the three, LightGBM was used for Lasso, Select K Best, PCA, and the pre-trained deep learning models. c . When using these models, we use stratiﬁed cross-validation with a k-fold of 3, tuning on the hyperparameter n estimators, which deﬁnes the complexity of the model. Summary ● Transfer learning techniques, speciﬁcally through ﬁne-tuning, performed best in binary and multi-accuracy for skin cancer image classiﬁcation. ● Fine-tuned transfer learning models showed lower AUC scores, possibly due to overﬁtting. ● Similar performance metrics can be obtained through multiple methods. Future Work ● A slightly higher cross-validation value can be used rather than 3. ● A more expanded range of hyperparameters can strengthen results. ● Using RGB images for the non-deep learning techniques can provide a more direct comparison to the deep learning techniques. References ● Bechelli, Solene, and Jerome Delhommelle. "Machine Learning and Deep Learning Algorithms for Skin Cancer Classiﬁcation from Dermoscopic Images. " Bioengineering 9.3 (2022): 97. ● Alkhushayni, Suboh, et al. "The Application of Diﬀering Machine Learning Algorithms and Their Related Performance in Detecting Skin Cancers and Melanomas. " Journal of Skin Cancer 2022 (2022) . ● Lynn, Nay Chi, and Zin Mar Kyu. "Segmentation and classiﬁcation of skin cancer melanoma from skin lesion images. " 2017 18th international conference on parallel and distributed computing, applications and tech

FAQs

How does the study reduce image dimensionality before classification?

All images are converted to grayscale and resized from 600x450 to 300x225. Dimensionality reduction is then applied via resizing percentages, Lasso, Select K Best, and PCA for the relevant approaches.

What feature selection and extraction techniques are compared?

Feature selection includes resizing-based reduction, Lasso with varying alpha values, and Select K Best on 74% sized images. Feature extraction includes PCA based on variance explained, and deep feature extraction using VGG16, Xception, and InceptionV3 before LightGBM classification.

Which approach performs best and how are binary vs multi-class results evaluated?

Fine-tuned transfer learning performs best for both binary and multi-class accuracy. Performance is measured using AUC for binary settings and accuracy for multi-class, computed with stratiﬁed 3-fold cross-validation and hyperparameter tuning of n_estimators.

Aug 6, 2026

Upload by Violet