Building Better Deep Learning Models Through Dataset Fusion: A Case Study in Skin Cancer Classification with...

Area of Science:

• Computer Science
• Medical Imaging
• Machine Learning

Background:

• Large, diverse image datasets are essential for robust machine learning model training.
• Managing and synthesizing large-scale datasets presents significant challenges for researchers.

Purpose of the Study:

• To introduce the Data Merger App for streamlining the creation of large-scale, diverse image datasets.
• To evaluate the impact of merged datasets on the performance of skin cancer classification models.

Main Methods:

• Developed a Data Merger App to identify common classes and combine diverse image datasets.
• Benchmarked Convolutional Neural Network (CNN) models (VGG16, ResNet50, MobileNetV3-small, DenseNet-161) and a Visual Transformer (ViT) for skin cancer classification.
• Compared model performance on single datasets versus enhanced hyperdatasets generated by the Data Merger App.

Main Results:

• Enhanced hyperdatasets significantly improved classification model accuracies for both training from scratch and Transfer Learning.
• The Visual Transformer (ViT) model achieved higher accuracies than CNNs, particularly with limited classes (91.87% for 9 classes) and on hyperdatasets (58% for 32 classes).

Conclusions:

• Data combination is vital for enhancing model generalization and improving the quality of research outcomes.
• The Data Merger App serves as a valuable tool for data scientists and researchers handling complex datasets.