MLCDL: A Critical Practice and Implementation of Multi-tissue Classification and Diagnosis Using Deep Learning...

Area of Science:

• Computer Science
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning (DL) has become a leading method for texture classification and tissue localization, surpassing traditional machine learning.
• Convolutional Neural Networks (CNNs) are a key component of DL, enabling advanced image analysis.
• Transfer learning enhances DL model performance by leveraging pre-trained networks.

Purpose of the Study:

• To introduce a new dataset for image-level texture classification.
• To evaluate the performance of an integrated transfer learning-based EfficientNet-B7 CNN for texture classification.
• To assess the framework's simplicity and potential for extracting prognostic information.

Main Methods:

• Development of a dataset comprising 381 images (150x150 pixels) for training, validation, and testing.
• Implementation of an EfficientNet-B7 deep convolutional neural network (CNN) model.
• Integration of transfer learning techniques to enhance classification accuracy.

Main Results:

• The EfficientNet-B7 model achieved high accuracy on the training dataset (89.33%).
• Validation and testing accuracies were 52.43% and 51.326%, respectively.
• Training losses were recorded as 0.2513 for training, 1.846 for validation, and 1.6137 for testing.

Conclusions:

• The proposed DL framework using EfficientNet-B7 demonstrates effectiveness for texture classification.
• The study highlights the straightforwardness of the framework's setup and execution.
• The technique shows potential for extracting more prognostic information compared to human analysis.