[Paper Review] ResNet101 and DAE for Enhance Quality and Classification Accuracy in Skin Cancer Imaging

Sibasish Dhibar|arXiv (Cornell University)|Mar 21, 2024

Cutaneous Melanoma Detection and ManagementMedicine3 citations

TL;DR

This paper proposes a hybrid deep learning framework combining ResNet101 and a Deep Autoencoder (DAE) to enhance skin lesion image quality and improve classification accuracy for skin cancer detection. The method leverages feature learning and denoising autoencoding to refine input images before classification, achieving 96.03% accuracy, 95.40% precision, 96.05% recall, 0.9576 F-measure, and 0.98 AUC on the ISIC-2018 dataset.

ABSTRACT

Skin cancer is a crucial health issue that requires timely detection for higher survival rates. Traditional computer vision techniques face challenges in addressing the advanced variability of skin lesion features, a gap partially bridged by convolutional neural networks (CNNs). To overcome the existing issues, we introduce an innovative convolutional ensemble network approach named deep autoencoder (DAE) with ResNet101. This method utilizes convolution-based deep neural networks for the detection of skin cancer. The ISIC-2018 public data taken from the source is used for experimental results, which demonstrate remarkable performance with the different in terms of performance metrics. The methods result in 96.03% of accuracy, 95.40 % of precision, 96.05% of recall, 0.9576 of F-measure, 0.98 of AUC.

Motivation & Objective

• To address the challenge of high variability in skin lesion features that hinders accurate skin cancer detection.
• To improve image quality of skin lesion scans for better feature extraction.
• To enhance classification performance using deep neural networks on real-world dermatoscopic images.
• To integrate a pre-trained ResNet101 with a Deep Autoencoder for joint feature learning and image reconstruction.
• To validate the proposed model on the ISIC-2018 benchmark dataset for clinical relevance.

Proposed method

• A Deep Autoencoder (DAE) is employed to reconstruct and enhance the quality of input skin lesion images by learning a compressed, noise-robust representation.
• The DAE is trained end-to-end on the ISIC-2018 dataset to learn hierarchical features and reduce image noise and artifacts.
• The enhanced images from the DAE are fed into a pre-trained ResNet101 backbone for feature extraction and classification.
• Transfer learning is applied by fine-tuning the ResNet101 classifier head on the ISIC-2018 dataset after DAE-based image preprocessing.
• The ensemble model is trained using cross-entropy loss with Adam optimization and evaluated using standard metrics including accuracy, precision, recall, F1-score, and AUC.

Experimental results

Research questions

• RQ1 Can a DAE-based image enhancement technique improve the quality of dermatoscopic images for better downstream classification?
• RQ2 How does combining ResNet101 with a DAE affect classification accuracy compared to using ResNet101 alone?
• RQ3 To what extent does the proposed method reduce the impact of image noise and variability in skin lesion appearance?
• RQ4 Does the joint use of autoencoding and residual learning lead to better generalization on the ISIC-2018 benchmark?
• RQ5 What is the performance of the model in terms of AUC, F1-score, and balanced metrics across skin cancer classes?

Key findings

• The proposed method achieved a classification accuracy of 96.03% on the ISIC-2018 dataset.
• The model demonstrated a precision of 95.40% and a recall of 96.05%, indicating strong balance between precision and sensitivity.
• The F-measure reached 0.9576, reflecting high harmonic mean of precision and recall.
• The area under the ROC curve (AUC) was 0.98, indicating excellent discriminative performance.
• The integration of DAE for image enhancement and ResNet101 for classification significantly outperformed baseline models in image quality and metric consistency.