Evaluating Deep Learning Models for Accurate Skin Lesion Segmentation
Discipline: Computer Sciences and Information Management
Session: 2
Room: 6 - Inman
LaVonne Wilson-Taylor - Fort Valley State University
Co-Author(s): Chunhua Dong, PhD, Fort Valley State University, Fort Valley, GA and Xiangyan Zeng, Phd, Fort Valley State University, Fort Valley, GA
Approximately one in five Americans will be diagnosed with skin cancer during their lifetime, making early detection critical for improving the five-year survival rate. Therefore, the automatic detection and classification of skin lesions plays a crucial role in a computer-aided diagnosis system, helping physicians make quicker and more accurate diagnoses. To facilitate this process, various image processing and machine learning techniques have been developed, including advanced deep learning algorithms based on convolutional neural networks (CNN). Automated diagnosis of skin lesions can be divided into two distinct phases: segmentation and classification. Segmentation is the identification and delineation of lesions, while classification is the prediction of the diagnostic category. In this study, funded by the National Science Foundation, HBCU-UP program, we conduct a comprehensive evaluation of deep learning (DL) models specifically for skin lesion segmentation. We evaluated 5 state-of-the-art deep learning models for this task: UNet, UNet++, ResUNet, ResUNet++, and DC-UNet. We used the HAM10000 dataset for our experiments, which consists of 10015 dermoscopic images of pigmented lesions from seven major diagnostic categories, along with its corresponding segmentation masks. The performance of each model was evaluated using accuracy, precision, recall, dice coefficient, jaccard index, and loss. Our study found that the ResUNet++ model outperformed the other models in every metric with a notable accuracy score of 96%. The model’s effectiveness can be attributed to its innovative architecture. It leverages Residual Blocks as a foundation and further strengthens its capabilities with the incorporation of Squeeze-and-Excite, Atrous Pyramidal Pooling, and Attention Blocks. These key enhancements improve the model’s ability to identify and emphasize the boarders of the lesion, therefore generating a more accurate segmented mask. Overall, this study compared the results of the DL models with hyperparameter tuning and identified ResUnet++ as the optimal model for segmentation. Further research will focus on evaluating and identifying the optimal model for skin lesion classification using the same dataset.
Keywords: skin lesion, skin cancer, convolutional neural network, image segmentation, deep learning
Funder Acknowledgement(s): NSF HPCU-UP, S-STEM and DoE MSEIP, and Google Inc.
Faculty Advisor: Xiangyan Zeng, Phd, xzeng@fvsu.edu
Role: I implemented the 5 deep learning models, trained them using a standardized dataset, and compared their results.