TLDR: A study by Nerma Kadric, Amila Akagic, and Medina Kapo investigates the use of Convolutional Neural Networks (CNNs), specifically ResNet-50 and EfficientNetV2L models, to improve the diagnostic accuracy of pigmented skin lesions. Using transfer learning on the DermaMNIST and an improved DermaMNIST-C dataset, the research found that EfficientNetV2L models, particularly with the higher-resolution DermaMNIST-C, achieved high accuracy and competitive performance in classifying various skin lesion types. The findings suggest CNNs can significantly enhance medical image analysis for better patient care, despite challenges like dataset imbalance.
Pigmented skin lesions, which are areas of increased melanin, can sometimes indicate serious conditions like melanoma, a significant contributor to skin cancer mortality. Early and accurate diagnosis is crucial for effective treatment and improving patient outcomes. This challenge is being addressed by advancements in artificial intelligence, particularly through the use of convolutional neural networks (CNNs).
A recent study explored the effectiveness of CNNs in improving the diagnostic accuracy of pigmented skin lesions. The researchers, Nerma Kadric, Amila Akagic, and Medina Kapo from the University of Sarajevo, focused on using the DermaMNIST dataset, which is part of the larger MedMNIST v2 collection. MedMNIST v2 was developed to support research in biomedical imaging, drawing inspiration from the well-known MNIST dataset used in early computer vision studies.
The DermaMNIST dataset is derived from HAM10000, a comprehensive collection of dermatoscopic images of common pigmented skin lesions. It categorizes images into seven different disease types, including melanoma and basal cell carcinoma. Initially, DermaMNIST images were resized to a small 28×28 pixels, and the dataset faced criticism for its organization and class imbalance. To address these issues, an improved version called DermaMNIST-C was introduced, featuring enhanced image resolution of 224×224 pixels and better organization.
The study specifically assessed two state-of-the-art deep learning models: ResNet-50 and EfficientNetV2L. Both models were pre-trained on ImageNet, a common practice known as transfer learning. Transfer learning allows models to leverage knowledge gained from training on a vast dataset of general images, which can then be fine-tuned for a specific task like medical image classification. The researchers froze the base model weights and added various layers on top, including flatten layers, batch normalization, dense layers, and dropout layers, to optimize performance for multi-class classification.
Two main experiments were conducted. The first experiment used the original DermaMNIST dataset (28×28 pixels). While both ResNet-50 and EfficientNetV2L models were tested, the results indicated issues with overfitting and suboptimal performance, suggesting that the original DermaMNIST dataset had limitations due to its small image size, class imbalance, and organization. The EfficientNetV2L model showed slight improvements over ResNet-50 in this initial test.
The second, more successful experiment utilized the improved DermaMNIST-C dataset (224×224 pixels). Here, the researchers evaluated five different model configurations, including a simple model trained from scratch and various EfficientNetV2L configurations. The EfficientNetV2L models demonstrated significant improvements. Notably, one configuration, Effv3 e, achieved the highest accuracy and recall among the tested models on DermaMNIST-C. This configuration simplified the classification structure by including a flatten layer, batch normalization, a dense layer with 64 neurons, another batch normalization layer, and a final dense layer for classification.
The study’s findings highlight that convolutional neural networks can significantly enhance diagnostic accuracy in biomedical image analysis. While a ResNet50 model from a previous study showed a slightly higher accuracy, the EfficientNetV2L model in this research demonstrated competitive and exceptional performance across multiple metrics, including precision, AUC (Area Under the Curve), and recall, which are crucial for medical diagnostics. The improved performance was largely attributed to the use of the higher-resolution and better-organized DermaMNIST-C dataset, along with the innovative modeling approach and feature extraction capabilities of EfficientNetV2L.
Also Read:
- AI’s Next Frontier: Vision Transformers Excel in Automated Pneumonia Detection from Chest X-Rays
- A New Unit-Based Approach Enhances Histopathology Tissue Segmentation
In conclusion, this research underscores the importance of sophisticated deep learning techniques in medical image analysis for classifying dermatological images. The study lays the groundwork for future developments in diagnostic tools that can significantly benefit patient care. Future work could involve expanding the dataset with generated images to further address class imbalance and integrating these classification models into real-world medical applications to aid in detecting malignant and benign skin changes. You can read the full research paper here: Improving Diagnostic Accuracy of Pigmented Skin Lesions With CNNs: A Case Study with DermaMNIST Dataset.
